OpenStack Design Guide

PREFACE

Introduction to the OpenStack Architecture Design
Guide
OpenStack is a leader in the cloud technology gold rush, as organizations of all
stripes discover the increased flexibility and speed to market that self-service cloud and
Infrastructure as a Service (IaaS) provides. To truly reap those benefits, however, the cloud
must be designed and architected properly.

A well-architected cloud provides a stable IT environment that offers easy access to needed
resources, usage-based expenses, extra capacity on demand, disaster recovery, and a
secure environment, but a well-architected cloud does not magically build itself. It requires
careful consideration of a multitude of factors, both technical and non-technical.
There is no single architecture that is "right" for an OpenStack cloud deployment. OpenStack
can be used for any number of different purposes, and each of them has its own particular
requirements and architectural peculiarities.
This book is designed to look at some of the most common uses for OpenStack clouds (and
even some that are less common, but provide a good example) and explain what issues need
to be considered and why, along with a wealth of knowledge and advice to help an
organization to design and build a well-architected OpenStack cloud that will fit its unique
requirements.

Intended Audience

This book has been written for architects and designers of OpenStack clouds. This book is
not intended for people who are deploying OpenStack. For a guide on deploying and
operating OpenStack, please refer to the Operations
Guide http://docs.openstack.org/openstack-ops.
The reader should have prior knowledge of cloud architecture and principles, experience in
enterprise system design, Linux and virtualization experience, and a basic understanding of
networking principles and protocols.

How this Book is Organized

This book has been organized into various chapters that help define the use cases
associated with making architectural choices related to an OpenStack cloud installation.
Each chapter is intended to stand alone to encourage individual chapter
readability, however each chapter is designed to contain useful information that may be
applicable in situations covered by other chapters. Cloud architects may use this book as a
comprehensive guide by reading all of the use cases, but it is also possible to review only the
chapters which pertain to a specific use case. When choosing to read specific use cases,
note that it may be necessary to read more than one section of the guide to formulate a
complete design for the cloud. The use cases covered in this guide include:
General purpose: A cloud built with common components that should address 80% of
common use cases.
Compute focused: A cloud designed to address compute intensive workloads such as
high performance computing (HPC).
Storage focused: A cloud focused on storage intensive workloads such as data analytics
with parallel filesystems.
Network focused: A cloud depending on high performance and reliable networking, such
as a content delivery network (CDN).
Multi-site: A cloud built with multiple sites available for application deployments for
geographical, reliability or data locality reasons.
Hybrid cloud: An architecture where multiple disparate clouds are connected either for
failover, hybrid cloud bursting, or availability.
Massively Scalable: An architecture that is intended for cloud service providers or other
extremely large installations.
A section titled Specialized Use Cases provides information on architectures that have not
previously been covered in the defined use cases.
Each chapter in the guide is then further broken down into the following sections:
Introduction: Provides an overview of the architectural use case.
User requirements: Defines the set of user considerations that typically come into play
for that use case.
Technical considerations: Covers the technical issues that must be accounted when
dealing with this use case.
Operational considerations: Covers the ongoing operational tasks associated with this
use case and architecture.
Architecture: Covers the overall architecture associated with the use case.

Prescriptive examples: Presents one or more scenarios where this architecture could be
deployed.
A Glossary covers the terms and phrases used in the book.

Why and How We Wrote this Book

The velocity at which OpenStack environments are moving from proof-of-concepts to
production deployments is leading to increasing questions and issues related to architecture
design considerations. By and large these considerations are not addressed in the existing
documentation, which typically focuses on the specifics of deployment and configuration
options or operational considerations, rather than the bigger picture.
We wrote this book to guide readers in designing an OpenStack architecture that meets the
needs of their organization. This guide concentrates on identifying important design
considerations for common cloud use cases and provides examples based on these design
guidelines. This guide does not aim to provide explicit instructions for installing and
configuring the cloud, but rather focuses on design principles as they relate to user
requirements as well as technical and operational considerations. For specific guidance with
installation and configuration there are a number of resources already available in the
OpenStack documentation that help in that area.
This book was written in a book sprint format, which is a facilitated, rapid development
production method for books. For more information, see the Book Sprints website
(www.booksprints.net).
This book was written in five days during July 2014 while exhausting the M&M, Mountain
Dew and healthy options supply, complete with juggling entertainment during lunches at
VMware's headquarters in Palo Alto. The event was also documented on Twitter using the
#OpenStackDesign hashtag. The Book Sprint was facilitated by Faith Bosworth and Adam
Hyde.
We would like to thank VMware for their generous hospitality, as well as our employers,
Cisco, Cloudscaling, Comcast, EMC, Mirantis, Rackspace, Red Hat, Verizon, and VMware, for
enabling us to contribute our time. We would especially like to think Anne Gentle and
Kenneth Hui for all of their shepherding and organization in making this happen.
The author team includes:
Kenneth Hui (EMC) @hui_kenneth
Alexandra Settle (Rackspace) @dewsday
Anthony Veiga (Comcast) @daaelar
Beth Cohen (Verizon) @bfcohen
Kevin Jackson (Rackspace) @itarchitectkev

Maish Saidel-Keesing (Cisco) @maishsk
Nick Chase (Mirantis) @NickChase
Scott Lowe (VMware) @scott_lowe
Sean Collins (Comcast) @sc68cal
Sean Winn (Cloudscaling) @seanmwinn
Sebastian Gutierrez (Red Hat) @gutseb
Stephen Gordon (Red Hat) @xsgordon
Vinny Valdez (Red Hat) @VinnyValdez

Methodology

Developing Cloud Use Cases

The magic of the cloud is that it can do anything. It is both robust and flexible, the best of
both worlds. Yes, the cloud is highly flexible and it can do almost anything, but to get the
most out of a cloud investment, it is important to define what the cloud will be used for by
creating and testing use cases. This is the chapter that describes the thought process behind
how to design a cloud architecture that best suits the intended use.

The diagram shows at a very abstract level the process for capturing requirements and
building use cases. Once a set of use cases has been defined, it can then be used to design
the cloud architecture.
Use case planning can seem counterintuitive. After all, it takes about five minutes to sign up
for a server with Amazon. Amazon does not know in advance what any given user
is planning on doing with it, right? Wrong. Amazon’s product management department
spends plenty of time figuring out exactly what would be attractive to their typical customer
and honing the service to deliver it. For the enterprise, the planning process is no different,
but instead of planning for an external paying customer, for example, the use could be for
internal application developers or a web portal. The following is a list of the high level
objectives that need to be incorporated into the thinking about creating a use case.
Overall business objectives

Develop clear definition of business goals and requirements
Increase project support and engagement with business, customers and end users
Technology
Coordinate the OpenStack architecture across the project and leverage OpenStack
community efforts more effectively
Architect for automation as much as possible to speed development and deployment
Use the appropriate tools for the development effort
Create better and more test metrics and test harnesses to support continuous and
integrated development, test processes and automation
Organization
Better messaging of management support of team efforts
Develop better cultural understanding of Open Source, cloud architectures, Agile
methodologies, continuous develepmont, test and integration, overall develo concepts in
general
As an example of how this works, consider a business goal of using the cloud for the
company’s E-commerce website. This goal means planning for applications that will support
thousands of sessions per second, variable workloads, and lots of complex and changing
data. By identifying the key metrics, such as number of concurrent transactions per second,
size of database, and so on, it is possible to then build a method for testing the
assumptions.
Develop functional user scenarios: Develop functional user scenarios that can be used to
develop test cases that can be used to measure overall project trajectory. If the
organization is not ready to commit to an application or applications that can be used to
develop user requirements, it needs to create requirements to build valid test harnesses and
develop useable metrics. Once the metrics are established, as requirements change, it is
easier to respond to the changes quickly without having to worry overly much about setting
the exact requirements in advance. Think of this as creating ways to configure the system,
rather than redesigning it every time there is a requirements change.
Limit cloud feature set: Create requirements that address the pain points, but do not
recreate the entire OpenStack tool suite. The requirement to build OpenStack, only better, is
self-defeating. It is important to limit scope creep by concentrating on developing a
platform that will address tool limitations for the requirements, but not recreating the
entire suite of tools. Work with technical product owners to establish critical features that

are needed for a successful cloud deployment.

Application Cloud Readiness

Although the cloud is designed to make things easier, it is important to realize that "using
cloud" is more than just firing up an instance and dropping an application on it. The "lift and
shift" approach works in certain situations, but there is a fundamental difference between
clouds and traditional bare-metal-based environments, or even traditional virtualized
environments.
In traditional environments, with traditional enterprise applications, the applications and the
servers that run on them are "pets". They're lovingly crafted and cared for, the servers have
names like Gandalf or Tardis, and if they get sick, someone nurses them back to health. All of
this is designed so that the application does not experience an outage.
In cloud environments, on the other hand, servers are more like cattle. There are thousands
of them, they get names like NY-1138-Q, and if they get sick, they get put down and a
sysadmin installs another one. Traditional applications that are unprepared for this kind of
environment, naturally will suffer outages, lost data, or worse.
There are other reasons to design applications with cloud in mind. Some are defensive, such
as the fact that applications cannot be certain of exactly where or on what hardware they
will be launched, they need to be flexible, or at least adaptable. Others are proactive. For
example, one of the advantages of using the cloud is scalability, so applications need to be
designed in such a way that they t can take advantage of those and other opportunities.

Determining whether an application is cloud-ready

There are several factors to take into consideration when looking at whether an application
is a good fit for the cloud.
Structure: A large, monolithic, single-tiered legacy application typically isn't a good fit for
the cloud. Efficiencies are gained when load can be spread over several instances, so that a
failure in one part of the system can be mitigated without affecting other parts of the
system, or so that scaling can take place where the app needs it.
Dependencies: Applications that depend on specific hardware -- such as a particular chipset
or an external device such as a fingerprint reader -- might not be a good fit for the cloud,
unless those dependencies are specifically addressed. Similarly, if an application depends on
an operating system or set of libraries that cannot be used in the cloud, or cannot be

virtualized, that is a problem.
Connectivity: Self-contained applications or those that depend on resources that are
not reachable by the cloud in question, will not run. In some situations, work around these
issues with custom network setup, but how well this works depends on the chosen cloud
environment.
Durability and Resilience: Despite the existence of SLAs, the one reality of the cloud is that
Things Break. Servers go down, network connections are disrupted, other tenants on a
server ramp up the load to make the server unusable. Any number of things can happen,
and an application that isn't built to withstand this kind of disruption isn't going to work
properly.

Designing for the cloud

Here are some guidelines to keep in mind when designing an application for the cloud:
Be a pessimist: Assume everything fails and design backwards. Love your chaos
monkey.
Put your eggs in multiple baskets: Leverage multiple providers, geographic regions
and availability zones to accommodate for local availability issues. Design for portability.
Think efficiency: Inefficient designs will not scale. Efficient designs become cheaper as
they scale. Kill off unneeded components or capacity.
Be paranoid: Design for defense in depth and zero tolerance by building in security at
every level and between every component. Trust no one.
But not too paranoid: Not every application needs the platinum solution. Architect for
different SLA’s, service tiers and security levels.
Manage the data: Data is usually the most inflexible and complex area of a cloud and
cloud integration architecture. Don’t short change the effort in analyzing and addressing
data needs.
Hands off: Leverage automation to increase consistency and quality and reduce
response times.
Divide and conquer: Pursue partitioning and parallelization wherever possible. Make
components as small and portable as possible. Use load balancing between layers.
Think elasticity: Increasing resources should result in a proportional increase in
performance and scalability. Decreasing resources should have the opposite effect.
Be dynamic: Enable dynamic configuration changes such as autoscaling, failure
recovery and resource discovery to adapt to changing environments, faults and
workload volumes.

Stay close: Reduce latency by moving highly interactive components and data near
each other.
Keep it loose: Loose coupling, service interfaces, separation of concerns, abstraction
and well defined API’s deliver flexibility.
Be cost aware: Autoscaling, data transmission, virtual software licenses, reserved
instances, and so on can rapidly increase monthly usage charges. Monitor usage closely.

GENERAL PURPOSE

Introduction

An OpenStack general purpose cloud is often considered a starting point for building a cloud
deployment. General purpose clouds, by their nature, balance the components and do not
emphasize (or heavily emphasize) any particular aspect of the overall computing
environment. The expectation is that the compute, network, and storage components will
be given equal weight in the design. General purpose clouds can be found in private, public,
and hybrid environments. They lend themselves to many different use cases but, since they
are homogeneous deployments, they are not suited to specialized environments or edge
case situations. Common uses to consider for a general purpose cloud could be, but are not
limited to, providing a simple database, a web application runtime environment, a
shared application development platform, or lab test bed. In other words, any use case that
would benefit from a scale-out rather than a scale-up approach is a good candidate for a
general purpose cloud architecture.
A general purpose cloud, by definition, is something that is designed to have a range of
potential uses or functions; not specialized for a specific use. General purpose architecture is
largely considered a scenario that would address 80% of the potential use cases. The
infrastructure, in itself, is a specific use case. It is also a good place to start the design
process. As the most basic cloud service model, general purpose clouds are designed to
be platforms suited for general purpose applications.
General purpose clouds are limited to the most basic components, but they can include
additional resources such as:
Virtual-machine disk image library
Raw block storage
File or object storage
Firewalls
Load balancers
IP addresses
Network overlays or virtual local area networks (VLANs)
Software bundles

User Requirements

The general purpose cloud is built following the Infrastructure-as-a-Service (IaaS) model; as
a platform best suited for use cases with simple requirements. The general purpose
cloud user requirements themselves are typically not complex. However, it is still important
to capture them even if the project has minimum business and technical
requirements such as a Proof of Concept (PoC) or a small lab platform.

User Considerations

These user considerations are written from the perspective of the organization that is
building the cloud, not from the perspective of the end-users who will consume cloud
services provided by this design.
Cost: Financial factors are a primary concern for any organization. Since general purpose
clouds are considered the baseline from which all other cloud architecture environments
derive, cost will commonly be an important criteria. This type of cloud, however, does
not always provide the most cost-effective environment for a specialized application or
situation. Unless razor-thin margins and costs have been mandated as a critical factor, cost
should not be the sole consideration when choosing or designing a general
purpose architecture.
Time to market: Another common business factor in building a general purpose cloud is
the ability to deliver a service or product more quickly and flexibly. In the modern hyperfast business world, being able to deliver a product in six months instead of two years is
often a major driving force behind the decision to build a general purpose cloud. General
purpose clouds allow users to self-provision and gain access to compute, network, and
storage resources on-demand thus decreasing time to market. It may potentially make
more sense to build a general purpose PoC as opposed to waiting to finalize the ultimate use
case for the system. The tradeoff with taking this approach is the risk that the general
purpose cloud is not optimized for the actual final workloads. The final decision on which
approach to take will be dependent on the specifics of the business objectives and time
frame for the project.
Revenue opportunity: The revenue opportunity for a given cloud will vary greatly based
on the intended use case of that particular cloud. Some general purpose clouds are built
for commercial customer facing products, but there are plenty of other reasons that might
make the general purpose cloud the right choice. A small cloud service provider (CSP) might
want to build a general purpose cloud rather than a massively scalable cloud because they
do not have the deep financial resources needed, or because they do not or will not know in
advance the purposes for which their customers are going to use the cloud. For some users,

the advantages cloud itself offers mean an enhancement of revenue opportunity. For
others, the fact that a general purpose cloud provides only baseline functionality will be a
disincentive for use, leading to a potential stagnation of potential revenue opportunities.

Legal Requirements

Many jurisdictions have legislative and regulatory requirements governing the storage and
management of data in cloud environments. Common areas of regulation include:
Data retention policies ensuring storage of persistent data and records management to
meet data archival requirements.
Data ownership policies governing the possession and responsibility for data.
Data sovereignty policies governing the storage of data in foreign countries or otherwise
separate jurisdictions.
Data compliance policies governing certain types of information need to reside in certain
locations due to regular issues - and more important cannot reside in other locations for the
same reason.
Examples of such legal frameworks include the data protection framework of the European
Union (http://ec.europa.eu/justice/data-protection/ ) and the requirements of the Financial
Industry Regulatory Authority (http://www.finra.org/Industry/Regulation/FINRARules/ ) in
the United States. Consult a local regulatory body for more information.

Technical Requirements

Technical cloud architecture requirements should be weighted against the
business requirements.
Performance: As a baseline product, general purpose clouds do not provide
optimized performance for any particular function. While a general purpose cloud should
provide enough performance to satisfy average user considerations, performance is not a
general purpose cloud customer driver.
No predefined usage model: The lack of a pre-defined usage model enables the user to
run a wide variety of applications without having to know the application requirements in
advance. This provides a degree of independence and flexibility that no other cloud
scenarios are able to provide.
On-demand and self-service application: By definition, a cloud provides end users with
the ability to self-provision computing power, storage, networks, and software in a simple
and flexible way. The user must be able to scale their resources up to a substantial level
without disrupting the underlying host operations. One of the benefits of using a general
purpose cloud architecture is the ability to start with limited resources and increase them

over time as the user demand grows.
Public cloud: For a company interested in building a commercial public cloud offering
based on OpenStack, the general purpose architecture model might be the best choice
because the designers are not going to know the purposes or workloads for which the end
users will use the cloud.
Internal consumption (private) cloud: Organizations need to determine if it makes the
most sense to create their own clouds internally. The main advantage of a private cloud is
that it allows the organization to maintain complete control over all the architecture and the
cloud components. One caution is to think about the possibility that users will want to
combine using the internal cloud with access to an external cloud. If that case is likely, it
might be worth exploring the possibility of taking a multi-cloud approach with regard to at
least some of the architectural elements. Designs that incorporate the use of multiple
clouds, such as a private cloud and a public cloud offering, are described in the "Multi-Cloud"
scenario.
Security: Security should be implemented according to asset, threat, and vulnerability risk
assessment matrices. For cloud domains that require increased computer security, network
security, or information security, general purpose cloud is not considered an appropriate
choice.

Technical Considerations

When designing a general purpose cloud, there is an implied requirement to design for all of
the base services generally associated with providing Infrastructure-as-a-Service: compute,
network and storage. Each of these services have different resource requirements. As a
result, it is important to make design decisions relating directly to the service currently under
design, while providing a balanced infrastructure that provides for all services.

Hardware Selection

When designing an OpenStack cloud as a general purpose cloud, the hardware selection
process can be lengthy and involved due to the sheer mass of services which need to be
designed and the unique characteristics and requirements of each service within the cloud.
Hardware designs need to be generated for each type of resource pool; specifically,
compute, network, and storage. In addition to the hardware designs, which affect the
resource nodes themselves, there are also a number of additional hardware decisions to be
made related to network architecture and facilities planning. These factors play heavily into
the overall architecture of an OpenStack cloud.

Designing Compute Resources

It is recommended to design compute resources as pools of resources which will be
addressed on-demand. When designing compute resource pools, a number of factors
impact your design decisions. For example, decisions related to processors, memory, and
storage within each hypervisor are just one element of designing compute resources. In
addition, it is necessary to decide whether compute resources will be provided in a single
pool or in multiple pools.
To design for the best use of available resources by applications running in the cloud, it is
recommended to design more than one compute resource pool. Each independent resource
pool should be designed to provide service for specific flavors of instances or groupings of
flavors. For the purpose of this book, "instance" refers to a virtual machines and
the operating system running on the virtual machine. Designing multiple resource pools
helps to ensure that, as instances are scheduled onto compute hypervisors, each
independent node's resources will be allocated in a way that makes the most efficient use of
available hardware. This is commonly referred to as bin packing.
Using a consistent hardware design among the nodes that are placed within a resource pool
also helps support bin packing. Hardware nodes selected for being a part of a compute

resource pool should share a common processor, memory, and storage layout. By choosing
a common hardware design, it becomes easier to deploy, support and maintain those nodes
throughout their lifecycle in the cloud.
OpenStack provides the ability to configure overcommit ratio--the ratio of virtual resources
available for allocation to physical resources present--for both CPU and memory. The
default CPU overcommit ratio is 16:1 and the default memory overcommit ratio is 1.5:1.
Determine the tuning of the overcommit ratios for both of these options during the design
phase, as this has a direct impact on the hardware layout of your compute nodes.
As an example, consider that a m1.small instance uses 1 vCPU, 20 GB of ephemeral storage
and 2,048 MB of RAM. When designing a hardware node as a compute resource pool to
service instances, take into consideration the number of processor cores available on the
node as well as the required disk and memory to service instances running at capacity. For a
server with 2 CPUs of 10 cores each, with hyperthreading turned on, the default CPU
overcommit ratio of 16:1 would allow for 640 (2 x 10 x 2 x 16) total m1.small instances. By
the same reasoning, using the default memory overcommit ratio of 1.5:1 you can determine
that the server will need at least 853GB (640 x 2,048 MB % 1.5) of RAM. When sizing nodes
for memory, it is also important to consider the additional memory required to service
operating system and service needs.
Processor selection is an extremely important consideration in hardware design, especially
when comparing the features and performance characteristics of different processors.
Some newly released processors include features specific to virtualized compute hosts
including hardware assisted virtualization and technology related to memory paging (also
known as EPT shadowing). These features have a tremendous positive impact on the
performance of virtual machines running in the cloud.
In addition to the impact on actual compute services, it is also important to consider the
compute requirements of resource nodes within the cloud. Resource nodes refers to nonhypervisor nodes providing controller, object storage, block storage, or networking services
in the cloud. The number of processor cores and threads has a direct correlation to the
number of worker threads which can be run on a resource node. It is important to ensure
sufficient compute capacity and memory is planned on resource nodes.
Workload profiles are unpredictable in a general purpose cloud, so it may be difficult to
design for every specific use case in mind. Additional compute resource pools can be added
to the cloud at a later time, so this unpredictability should not be a problem. In some cases,

the demand on certain instance types or flavors may not justify an individual hardware
design. In either of these cases, start by providing hardware designs which will be capable of
servicing the most common instance requests first, looking to add additional hardware
designs to the overall architecture in the form of new hardware node designs and resource
pools as they become justified at a later time.

Designing Network Resources

An OpenStack cloud traditionally has multiple network segments, each of which provides
access to resources within the cloud to both operators and tenants. In addition, the network
services themselves also require network communication paths which should also be
separated from the other networks. When designing network services for a general purpose
cloud, it is recommended to plan for either a physical or logical separation of network
segments which will be used by operators and tenants. It is further suggested to create an
additional network segment for access to internal services such as the message bus and
database used by the various cloud services. Segregating these services onto separate
networks helps to protect sensitive data and also protects against unauthorized access to
services.
Based on the requirements of instances being serviced in the cloud, the next design choice
which will affect your design is the choice of network service which will be used to service
instances in the cloud. The choice between nova-network, as a part of OpenStack Compute
Service, and Neutron, the OpenStack Networking Service, has tremendous implications and
will have a huge impact on the architecture and design of the cloud network infrastructure.
The nova-network service is primarily a layer 2 networking service which has two main
modes in which it will function. The difference between the two modes in nova-network
pertain to whether or not nova-network uses VLANs. When using nova-network in a flat
network mode, all network hardware nodes and devices throughout the cloud are
connected to a single layer 2 network segment which provides access to application data.
When the network devices in the cloud support segmentation using VLANs, nova-network
can operate in the second mode. In this design model, each tenant within the cloud is
assigned a network subnet which is mapped to a VLAN on the physical network. It is
especially important to remember the maximum number of 4096 VLANs which can be used
within a spanning tree domain. These limitations place hard limits on the amount of growth
possible within the data center. When designing a general purpose cloud intended to
support multiple tenants, it is especially recommended to use nova-network with VLANs,

and not in flat network mode.
Another consideration regarding network is the fact that nova-network is entirely managed
by the cloud operator; tenants do not have control over network resources. If tenants
require the ability to manage and create network resources such as network segments and
subnets, it will be necessary to install the OpenStack Networking Service to provide network
access to instances.
The OpenStack Networking Service is a first class networking service that gives full control
over creation of virtual network resources to tenants. This is often accomplished in the form
of tunneling protocols which will establish encapsulated communication paths over existing
network infrastructure in order to segment tenant traffic. These methods vary depending on
the specific implementation, but some of the more common methods include tunneling
over GRE, encapsulating with VXLAN, and VLAN tags.
Initially, it is suggested to design at least three network segments, the first of which will be
used for access to the cloud’s REST APIs by tenants and operators. This is generally referred
to as a public network. In most cases, the controller nodes and swift proxies within the cloud
will be the only devices necessary to connect to this network segment. In some cases, this
network might also be serviced by hardware load balancers and other network devices.
The next segment is used by cloud administrators to manage hardware resources and is also
used by configuration management tools when deploying software and services onto new
hardware. In some cases, this network segment might also be used for internal services,
including the message bus and database services, to communicate with each other. Due to
the highly secure nature of this network segment, it may be desirable to secure this network
from unauthorized access. This network will likely need to communicate with every
hardware node within the cloud.
The last network segment is used by applications and consumers to provide access to the
physical network and also for users accessing applications running within the cloud. This
network is generally segregated from the one used to access the cloud APIs and is not
capable of communicating directly with the hardware resources in the cloud. Compute
resource nodes will need to communicate on this network segment, as will any network
gateway services which allow application data to access the physical network outside of the
cloud.

Designing Storage Resources

OpenStack has two independent storage services to consider, each with its own specific
design requirements and goals. In addition to services which provide storage as their primary
function, there are additional design considerations with regard to compute and controller
nodes which will affect the overall cloud architecture.

Designing OpenStack Object Storage

When designing hardware resources for OpenStack Object Storage, the primary goal is to
maximize the amount of storage in each resource node while also ensuring that the cost per
terabyte is kept to a minimum. This often involves utilizing servers which can hold a large
number of spinning disks. Whether choosing to use 2U server form factors with directly
attached storage or an external chassis that holds a larger number of drives, the main goal is
to maximize the storage available in each node.
It is not recommended to invest in enterprise class drives for an OpenStack Object Storage
cluster. The consistency and partition tolerance characteristics of OpenStack Object Storage
will ensure that data stays up to date and survives hardware faults without the use of any
specialized data replication devices.
A great benefit of OpenStack Object Storage is the ability to mix and match drives by utilizing
weighting within the swift ring. When designing your swift storage cluster, it is
recommended to make use of the most cost effective storage solution available at the time.
Many server chassis on the market can hold 60 or more drives in 4U of rack space, therefore
it is recommended to maximize the amount of storage per rack unit at the best cost per
terabyte. Furthermore, the use of RAID controllers is not recommended in an object storage
node.
In order to achieve this durability and availability of data stored as objects, it is important to
design object storage resource pools in a way that provides the suggested availability that
the service can provide. Beyond designing at the hardware node level, it is important to
consider rack-level and zone-level designs to accommodate the number of replicas
configured to be stored in the Object Storage service (the default number of replicas is
three). Each replica of data should exist in its own availability zone with its own power,
cooling, and network resources available to service that specific zone.
Object storage nodes should be designed so that the number of requests does not hinder
the performance of the cluster. The object storage service is a chatty protocol, therefore
making use of multiple processors that have higher core counts will ensure the IO requests
do not inundate the server.

Designing OpenStack Block Storage

When designing OpenStack Block Storage resource nodes, it is helpful to understand the
workloads and requirements that will drive the use of block storage in the cloud. In a general
purpose cloud these use patterns are often unknown. It is recommended to design block
storage pools so that tenants can choose the appropriate storage solution for their
applications. By creating multiple storage pools of different types, in conjunction with
configuring an advanced storage scheduler for the block storage service, it is possible to
provide tenants with a large catalog of storage services with a variety of performance levels
and redundancy options.
In addition to directly attached storage populated in servers, block storage can also take
advantage of a number of enterprise storage solutions. These are addressed via a plugin
driver developed by the hardware vendor. A large number of enterprise storage plugin
drivers ship out-of-the-box with OpenStack Block Storage (and many more available via
third party channels). While a general purpose cloud would likely use directly attached
storage in the majority of block storage nodes, it may also be necessary to provide
additional levels of service to tenants which can only be provided by enterprise class storage
solutions.
The determination to use a RAID controller card in block storage nodes is impacted primarily
by the redundancy and availability requirements of the application. Applications which have
a higher demand on input-output per second (IOPS) will influence both the choice to use a
RAID controller and the level of RAID configured on the volume. Where performance is a
consideration, it is suggested to make use of higher performing RAID volumes. In contrast,
where redundancy of block storage volumes is more important it is recommended to make
use of a redundant RAID configuration such as RAID 5 or RAID 6. Some specialized features,
such as automated replication of block storage volumes, may require the use of third-party
plugins and enterprise block storage solutions in order to provide the high demand on
storage. Furthermore, where extreme performance is a requirement it may also be
necessary to make use of high speed SSD disk drives' high performing flash storage
solutions.

Software Selection

The software selection process can play a large role in the architecture of a general purpose
cloud. Choice of operating system, selection of OpenStack software components, choice of
hypervisor and selection of supplemental software will have a large impact on the design of
the cloud.

Operating System

Operating system (OS) selection plays a large role in the design and architecture of a cloud.
There are a number of OSes which have native support for OpenStack including Ubuntu, Red
Hat Enterprise Linux (RHEL), CentOS, and SUSE Linux Enterprise Server (SLES). "Native
support" in this context means that the distribution provides distribution-native packages
by which to install OpenStack in their repositories. Note that "native support" is not a
constraint on the choice of OS; users are free to choose just about any Linux distribution (or
even Microsoft Windows) and install OpenStack directly from source (or compile their own
packages). However, the reality is that many organizations will prefer to install OpenStack
from distribution-supplied packages or repositories (although using the distribution
vendor's OpenStack packages might be a requirement for support).
OS selection also directly influences hypervisor selection. A cloud architect who selects
Ubuntu or RHEL has some flexibility in hypervisor; KVM, Xen, and LXC are supported
virtualization methods available under OpenStack Compute (Nova) on these Linux
distributions. A cloud architect who selects Hyper-V, on the other hand, is limited to
Windows Server. Similarly, a cloud architect who selects XenServer is limited to the CentOSbased dom0 operating system provided with XenServer.
The primary factors that play into OS/hypervisor selection include:
User requirements: The selection of OS/hypervisor combination first and foremost needs
to support the user requirements.
Support: The selected OS/hypervisor combination needs to be supported by OpenStack.
Interoperability: The OS/hypervisor needs to be interoperable with other features and
services in the OpenStack design in order to meet the user requirements.

Hypervisor

OpenStack supports a wide variety of hypervisors, one or more of which can be used in a
single cloud. These hypervisors include:
KVM (and Qemu)
XCP/XenServer
vSphere (vCenter and ESXi)
Hyper-V
LXC
Docker
Bare-metal

A complete list of supported hypervisors and their capabilities can be found
at https://wiki.openstack.org/wiki/HypervisorSupportMatrix.
General purpose clouds should make use of hypervisors that support the most general
purpose use cases, such as KVM and Xen. More specific hypervisors should then be chosen
to account for specific functionality or a supported feature requirement. In some cases,
there may also be a mandated requirement to run software on a certified hypervisor
including solutions from VMware, Microsoft, and Citrix.
The features offered through the OpenStack cloud platform determine the best choice of
a hypervisor. As an example, for a general purpose cloud that predominantly supports a
Microsoft-based migration, or is managed by staff that has a particular skill for managing
certain hypervisors and operating systems, Hyper-V might be the best available choice.
While the decision to use Hyper-V does not limit the ability to run alternative operating
systems, be mindful of those that are deemed supported. Each different hypervisor also has
their own hardware requirements which may affect the decisions around designing a
general purpose cloud. For example, to utilize the live migration feature of VMware,
vMotion, this requires an installation of vCenter/vSphere and the use of the ESXi hypervisor,
which increases the infrastructure requirements.
In a mixed hypervisor environment, specific aggregates of compute resources, each with
defined capabilities, enable workloads to utilize software and hardware specific to thier
particular requirements. This functionality can be exposed explicitly to the end user, or
accessed through defined metadata within a particular flavor of an instance.

OpenStack Components

A general purpose OpenStack cloud design should incorporate the core OpenStack services
to provide a wide range of services to end-users. The OpenStack core services
recommended in a general purpose cloud are:
OpenStack Compute (Nova)
OpenStack Networking (Neutron)
OpenStack Image Service (Glance)
OpenStack Identity Service (Keystone)
OpenStack Dashboard (Horizon)
OpenStack Telemetry (Ceilometer)
A general purpose cloud may also include OpenStack Object Storage (Swift). OpenStack
Block Storage (Cinder) may be selected to provide persistent storage to applications and

instances although, depending on the use case, this could be optional.

Supplemental Software

A general purpose OpenStack deployment consists of more than just OpenStack-specific
components. A typical deployment involves services that provide supporting functionality,
including databases and message queues, and may also involve software to provide
high availabilty of the OpenStack environment. Design decisions around the underlying
message queue might affect the required number of controller services, as well as the
technology to provide highly resilient database functionality, such as MariaDB with Galera. In
such a scenario, replication of services relies on quorum. Therefore, the underlying database
nodes, for example, should consist of at least 3 nodes to account for the recovery of a failed
Galera node. When increasing the number of nodes to support a feature of the software,
consideration of rack space and switch port density becomes important.
Where many general purpose deployments use hardware load balancers to provide highly
available API access and SSL termination, software solutions, for example HAProxy, can also
be considered. It is vital to ensure that such software implementations are also made highly
available. This high availability can be achieved by using software such as Keepalived or
Pacemaker with Corosync. Pacemaker and Corosync can provide Active-Active or ActivePassive highly available configuration depending on the specific service in the OpenStack
environment. Using this software can affect the design as it assumes at least a 2-node
controller infrastructure where one of those nodes may be running certain services in
standby mode.
Memcached is a distributed memory object caching system, and Redis is a key-value store.
Both are usually deployed on general purpose clouds to assist in alleviating load to the
Identity service. The memcached service caches tokens, and due to its distributed nature it
can help alleviate some bottlenecks to the underlying authentication system. Using
memcached or Redis does not affect the overall design of your architecture as they tend to
be deployed onto the infrastructure nodes providing the OpenStack services.

Performance

Performance of an OpenStack deployment is dependent on a number of factors related to
the infrastructure and controller services. The user requirements can be split into general
network performance, performance of compute resources, and performance of storage
systems.

Controller Infrastructure

The Controller infrastructure nodes provide management services to the end-user as well
as providing services internally for the operating of the cloud. The Controllers typically run
message queuing services that carry system messages between each service. Performance
issues related to the message bus would lead to delays in sending that message to where it
needs to go. The result of this condition would be delays in operation functions such as
spinning up and deleting instances, provisioning new storage volumes and managing
network resources. Such delays could adversely affect an application’s ability to react to
certain conditions, especially when using auto-scaling features. It is important to properly
design the hardware used to run the controller infrastructure as outlined above in the
Hardware Selection section.
Performance of the controller services is not just limited to processing power, but
restrictions may emerge in serving concurrent users. Ensure that the APIs and Horizon
services are load tested to ensure that you are able to serve your customers. Particular
attention should be made to the OpenStack Identity Service (Keystone), which provides the
authentication and authorization for all services, both internally to OpenStack itself and to
end-users. This service can lead to a degradation of overall performance if this is not sized
appropriately.

Network Performance

In a general purpose OpenStack cloud, the requirements of the network help determine its
performance capabilities. For example, small deployments may employ 1 Gibabit Ethernet
(GbE) networking, whereas larger installations serving multiple departments or many users
would be better architected with 10 GbE networking. The performance of the running
instances will be limited by these speeds. It is possible to design OpenStack environments
that run a mix of networking capabilities. By utilizing the different interface speeds, the users
of the OpenStack environment can choose networks that are fit for their purpose. For
example, web application instances may run on a public network presented through
OpenStack Networking that has 1 GbE capability, whereas the backend database uses an
OpenStack Networking network that has 10 GbE capability to replicate its data or, in some
cases, the design may incorporate link aggregation for greater throughput.
Network performance can be boosted considerably by implementing hardware load
balancers to provide front-end service to the cloud APIs. The hardware load balancers also
perform SSL termination if that is a requirement of your environment. When implementing
SSL offloading, it is important to understand the SSL offloading capabilities of the devices

selected.

Compute Host

The choice of hardware specifications used in compute nodes including CPU, memory and
disk type directly affects the performance of the instances. Other factors which can directly
affect performance include tunable parameters within the OpenStack services, for
example the overcommit ratio applied to resources. The defaults in OpenStack Compute set
a 16:1 over-commit of the CPU and 1.5 over-commit of the memory. Running at such high
ratios leads to an increase in "noisy-neighbor" activity. Care must be taken when sizing your
Compute environment to avoid this scenario. For running general purpose OpenStack
environments it is possible to keep to the defaults, but make sure to monitor your
environment as usage increases.

Storage Performance

When considering performance of OpenStack Block Storage, hardware and architecture
choice is important. Block Storage can use enterprise backend systems such as NetApp or
EMC, use scale out storage such as GlusterFS and Ceph, or simply use the capabilities of
directly attached storage in the nodes themselves. Block Storage may be deployed so that
traffic traverses the host network, which could affect, and be adversely affected by, the
front-side API traffic performance. As such, consider using a dedicated data storage
network with dedicated interfaces on the Controller and Compute hosts.
When considering performance of OpenStack Object Storage, a number of design choices
will affect performance. A user’s access to the Object Storage is through the proxy services,
which typically sit behind hardware load balancers. By the very nature of a highly resilient
storage system, replication of the data would affect performance of the overall system. In
this case, 10 GbE (or better) networking is recommended throughout the storage network
architecture.

Availability

In OpenStack, the infrastructure is integral to providing services and should always be
available, especially when operating with SLAs. Ensuring network availability is
accomplished by designing the network architecture so that no single point of failure exists.
A consideration of the number of switches, routes and redundancies of power should be
factored into core infrastructure, as well as the associated bonding of networks to provide
diverse routes to your highly available switch infrastructure.

The OpenStack services themselves should be deployed across multiple servers that do not
represent a single point of failure. Ensuring API availability can be achieved by placing these
services behind highly available load balancers that have multiple OpenStack servers as
members.
OpenStack lends itself to deployment in a highly available manner where it is expected that
at least 2 servers be utilized. These can run all the services involved from the message
queuing service, for example RabbitMQ or QPID, and an appropriately deployed database
service such as MySQL or MariaDB. As services in the cloud are scaled out, backend services
will need to scale too. Monitoring and reporting on server utilization and response times, as
well as load testing your systems, will help determine scale out decisions.
Care must be taken when deciding network functionality. Currently, OpenStack supports
both the legacy Nova-network system and the newer, extensible OpenStack Networking.
Both have their pros and cons when it comes to providing highly available access. Novanetwork, which provides networking access maintained in the OpenStack Compute code,
provides a feature that removes a single point of failure when it comes to routing, and this
feature is currently missing in OpenStack Networking. The effect of Nova network’s MultiHost functionality restricts failure domains to the host running that instance.
On the other hand, when using OpenStack Networking, the OpenStack controller servers or
separate OpenStack Networking hosts handle routing. For a deployment that
requires features available in only OpenStack Networking, it is possible to remove this
restriction by using third party software that helps maintain highly available L3 routes. Doing
so allows for common APIs to control network hardware, or to provide complex multi-tier
web applications in a secure manner. It is also possible to completely remove routing
from OpenStack Networking, and instead rely on hardware routing capabilities. In this case,
the switching infrastructure must support L3 routing.
OpenStack Networking (Neutron) and Nova Network both have their advantages and
disadvantages. They are both valid and supported options that fit different use cases as
described in the following table.
OpenStack Networking vs Nova Network
Nova Network
Simple, single agent
More mature, established
Flat or VLAN
No plugin support

OpenStack Networking
Complex, multiple agents
Newer, maturing
Flat, VLAN, Overlays, L2-L3, SDN
Plugin support for 3rd parties

Scales well
No multi-tier topologies

Scaling requires 3rd party plugins
Multi-tier topologies

Ensure your deployment has adequate back-up capabilities. As an example, in a deployment
that has two infrastructure controller nodes, the design should include controller
availability. In the event of the loss of a single controller, cloud services will run from a single
controller in the event of failure. Where the design has higher availability requirements, it is
important to meet those requirements by designing the proper redundancy and availability
of controller nodes.
Application design must also be factored into the capabilities of the underlying cloud
infrastructure. If the compute hosts do not provide a seamless live migration capability, then
it must be expected that when a compute host fails, that instance and any data local to that
instance will be deleted. Conversely, when providing an expectation to users that instances
have a high-level of uptime guarantees, the infrastructure must be deployed in a way
that eliminates any single point of failure when a compute host disappears. This may include
utilizing shared file systems on enterprise storage or OpenStack Block storage to provide a
level of guarantee to match service features.
For more information on HA in OpenStack, see the OpenStack High Availablility Guide found
at http://docs.openstack.org/high-availability-guide.

Security

A security domain comprises users, applications, servers or networks that share common
trust requirements and expectations within a system. Typically they have the same
authentication and authorization requirements and users.
These security domains are:
Public
Guest
Management
Data
These security domains can be mapped to an OpenStack deployment individually, or
combined. For example, some deployment topologies combine both guest and data
domains onto one physical network, whereas in other cases these networks are physically
separated. In each case, the cloud operator should be aware of the appropriate security
concerns. Security domains should be mapped out against your specific OpenStack

deployment topology. The domains and their trust requirements depend upon whether the
cloud instance is public, private, or hybrid.
The public security domain is an entirely untrusted area of the cloud infrastructure. It can
refer to the Internet as a whole or simply to networks over which you have no
authority. This domain should always be considered untrusted.
Typically used for compute instance-to-instance traffic, the guest security domain handles
compute data generated by instances on the cloud but not services that support the
operation of the cloud, such as API calls. Public cloud providers and private cloud providers
who do not have stringent controls on instance use or who allow unrestricted internet
access to instances should consider this domain to be untrusted. Private cloud providers
may want to consider this network as internal and therefore trusted only if they have
controls in place to assert that they trust instances and all their tenants.
The management security domain is where services interact. Sometimes referred to as the
"control plane", the networks in this domain transport confidential data such as
configuration parameters, usernames, and passwords. In most deployments this domain is
considered trusted.
The data security domain is concerned primarily with information pertaining to the storage
services within OpenStack. Much of the data that crosses this network has high integrity and
confidentiality requirements and, depending on the type of deployment, may also
have strong availability requirements. The trust level of this network is heavily dependent
on other deployment decisions.
When deploying OpenStack in an enterprise as a private cloud it is usually behind the
firewall and within the trusted network alongside existing systems. Users of the cloud are,
traditionally, employees that are bound by the security requirements set forth by the
company. This tends to push most of the security domains towards a more trusted model.
However, when deploying OpenStack in a public facing role, no assumptions can be made
and the attack vectors significantly increase. For example, the API endpoints, along with the
software behind them, become vulnerable to bad actors wanting to gain unauthorized
access or prevent access to services, which could lead to loss of data, functionality, and
reputation. These services must be protected against through auditing and appropriate
filtering.
Consideration must be taken when managing the users of the system for both public

and private clouds. The identity service allows for LDAP to be part of the authentication
process. Including such systems in an OpenStack deployment may ease user management if
integrating into existing systems.
It's important to understand that user authentication requests include sensitive information
including usernames, passwords and authentication tokens. For this reason, placing the API
services behind hardware that performs SSL termination is strongly recommended.
For more information OpenStack Security, see the OpenStack Security Guide,
at http://docs.openstack.org/security-guide/.

Operational Considerations

Many operational factors will affect general purpose cloud design choices. In larger
installations, it is not uncommon for operations staff to be tasked with maintaining cloud
environments. This differs from the operations staff that is responsible for building or
designing the infrastructure. It is important to include the operations function in the
planning and design phases of the build out.

Service Level Agreements

Service Level Agreements (SLAs) are contractual obligations that provide assurances for
service availability. SLAs define levels of availability that drive the technical design, often
with penalties for not meeting the contractual obligations. The strictness of the SLA dictates
the level of redundancy and resiliency in the OpenStack cloud design. Knowing when and
where to implement redundancy and HA is directly affected by expectations set by the
terms of the SLA. Some of the SLA terms that will affect the design include:
Guarantees for API availability imply multiple infrastructure services combined with
highly available load balancers.
Network uptime guarantees will affect the switch design and might
require redundant switching and power.
Network security policies requirements need to be factored in to deployments.

Support and Maintainability

OpenStack cloud management requires operations staff to be able to understand and
comprehend design architecture content on some level. The level of skills and the level of
separation of the operations and engineering staff are dependent on the size and purpose of
the installation. A large cloud service provider or a telecom provider is more likely to be
managed by a specially trained, dedicated operations organization. A smaller
implementation is more likely to rely on a smaller support staff that might need to take on
the combined engineering, design and operations functions.
Furthermore, maintaining OpenStack installations requires a variety of technical skills. Some
of these skills may include the ability to debug Python log output to a basic level and an
understanding of networking concepts.
Consider incorporating features into the architecture and design that reduce the operations
burden. This is accomplished by automating some of the operations functions. In some
cases it may be beneficial to use a third party management company with special expertise

in managing OpenStack deployments.

Monitoring

Like any other infrastructure deployment, OpenStack clouds need an appropriate
monitoring platform to ensure any errors are caught and managed appropriately. Consider
leveraging any existing monitoring system to see if it will be able to effectively monitor an
OpenStack environment. While there are many aspects that need to be monitored, specific
metrics that are critically important to capture include image disk utilization, or response
time to the Nova API.

Downtime

No matter how robust the architecture is, at some point components will fail. Designing for
high availability (HA) can have significant cost ramifications, therefore the resiliency of the
overall system and the individual components is going to be dictated by the requirements of
the SLA. Downtime planning includes creating processes and architectures that support
planned (maintenance) and unplanned (system faults) downtime.
An example of an operational consideration is the recovery of a failed compute host. This
might mean requiring the restoration of instances from a snapshot or respawning an
instance on another available compute host. This could have consequences on the overall
application design. A general purpose cloud should not need to provide an ability to migrate
instances from one host to another. If the expectation is that the application will be designed
to tolerate failure, additional considerations need to be made around supporting instance
migration. In this scenario, extra supporting services, including shared storage attached to
compute hosts, might need to be deployed.

Capacity Planning

Capacity planning for future growth is a critically important and often overlooked
consideration. Capacity constraints in a general purpose cloud environment
include compute and storage limits. There is a relationship between the size of the compute
environment and the supporting OpenStack infrastructure controller nodes required to
support it. As the size of the supporting compute environment increases, the network traffic
and messages will increase which will add load to the controller or networking nodes. While
no hard and fast rule exists, effective monitoring of the environment will help with capacity
decisions on when to scale the back-end infrastructure as part of the scaling of the compute
resources.

Adding extra compute capacity to an OpenStack cloud is a horizontally scaling process as
consistently configured compute nodes automatically attach to an OpenStack cloud. Be
mindful of any additional work that is needed to place the nodes into appropriate availability
zones and host aggregates. Make sure to use identical or functionally compatible CPUs
when adding additional compute nodes to the environment otherwise live migration
features will break. Scaling out compute hosts will directly affect network and other
datacenter resources so it will be necessary to add rack capacity or network switches.
Another option is to assess the average workloads and increase the number of instances
that can run within the compute environment by adjusting the overcommit ratio. While only
appropriate in some environments, it's important to remember that changing the CPU
overcommit ratio can have a detrimental effect and cause a potential increase in noisy
neighbor. The added risk of increasing the overcommit ratio is more instances will fail when
a compute host fails.
Compute host components can also be upgraded to account for increases in demand; this
is known as vertical scaling. Upgrading CPUs with more cores, or increasing the overall
server memory, can add extra needed capacity depending on whether the running
applications are more CPU intensive or memory intensive.
Insufficient disk capacity could also have a negative effect on overall performance including
CPU and memory usage. Depending on the back-end architecture of the OpenStack Block
Storage layer, capacity might include adding disk shelves to enterprise storage
systems or installing additional block storage nodes. It may also be necessary to upgrade
directly attached storage installed in compute hosts or add capacity to the shared storage to
provide additional ephemeral storage to instances.
For a deeper discussion on many of these topics, refer to the OpenStack Operations Guide at
http://docs.openstack.org/ops.

Architecture

Hardware Selection

Hardware selection involves three key areas:
Compute
Network
Storage
For each of these areas, the selection of hardware for a general purpose OpenStack cloud
must reflect the fact that a the cloud has no pre-defined usage model. This means that
there will be a wide variety of applications running on this cloud that will have varying
resource usage requirements. Some applications will be RAM-intensive, some applications
will be CPU-intensive, while others will be storage-intensive. Therefore, choosing
hardware for a general purpose OpenStack cloud must provide balanced access to all major
resources.

Selecting Compute (Server) Hardware

Certain hardware form factors may be better suited for use in a general purpose OpenStack
cloud because of the need for an equal or nearly equal balance of resources. Server
hardware for a general purpose OpenStack architecture design must provide an equal or
nearly equal balance of compute capacity (RAM and CPU), network capacity (number and
speed of links), and storage capacity (gigabytes or terabytes as well as I-O Operations Per
Second (IOPS).
Server hardware is evaluated around four conflicting dimensions:
Server density: A measure of how many servers can fit into a given measure of physical
space, such as a rack unit [U].
Resource capacity: The number of CPU cores, how much RAM, or how much storage a
given server will deliver.
Expandability: The number of additional resources that can be added to a server before it
has reached its limit.
Cost: The relative purchase price of the hardware weighted against the level of design
effort needed to build the system.
Increasing server density means sacrificing resource capacity or expandability, however,
increasing resource capacity and expandability increases cost and decreases server density.
As a result, determining the best server hardware for a general purpose OpenStack
architecture means understanding how choice of form factor will impact the rest of the

design.
Blade servers typically support dual-socket multi-core CPUs, which is the configuration
generally considered to be the "sweet spot" for a general purpose cloud deployment. Blades
also offer outstanding density. As an example, both HP BladeSystem and Dell PowerEdge
M1000e support up to 16 servers in only 10 rack units. However, the blade servers
themselves often have limited storage and networking capacity. Additionally, the
expandability of many blade servers can be limited.
1U rack-mounted servers occupy only a single rack unit. Their benefits include high
density, support for dual-socket multi-core CPUs, and support for reasonable RAM
amounts. This form factor offers limited storage capacity, limited network capacity, and
limited expandability.
2U rack-mounted servers offer the expanded storage and networking capacity that 1U
servers tend to lack, but with a corresponding decrease in server density (half the density
offered by 1U rack-mounted servers).
Larger rack-mounted servers, such as 4U servers, will tend to offer even greater CPU
capacity, often supporting four or even eight CPU sockets. These servers often have much
greater expandability so will provide the best option for upgradability. This means, however,
that the servers have a much lower server density and a much greater hardware cost.
"Sled servers" are rack-mounted servers that support multiple independent servers in a
single 2U or 3U enclosure. This form factor offers increased density over typical 1U-2U
rack-mounted servers but tends to suffer from limitations in the amount of storage or
network capacity each individual server supports.
Given the wide selection of hardware and general user requirements, the best form factor
for the server hardware supporting a general purpose OpenStack cloud is driven by outside
business and cost factors. No single reference architecture will apply to all implementations;
the decision must flow out of the user requirements, technical considerations, and
operational considerations. Here are some of the key factors that influence the selection of
server hardware:
Instance density: Sizing is an important consideration for a general purpose OpenStack
cloud. The expected or anticipated number of instances that each hypervisor can host is a
common metric used in sizing the deployment. The selected server hardware needs to
support the expected or anticipated instance density.
Host density: Physical data centers have limited physical space, power, and cooling. The
number of hosts (or hypervisors) that can be fitted into a given metric (rack, rack unit, or
floor tile) is another important method of sizing. Floor weight is an often overlooked
consideration. The data center floor must be able to support the weight of the proposed

number of hosts within a rack or set of racks. These factors need to be applied as part of the
host density calculation and server hardware selection.
Power density: Data centers have a specified amount of power fed to a given rack or set
of racks. Older data centers may have a power density as power as low as 20 AMPs per rack,
while more recent data centers can be architected to support power densities as high as 120
AMP per rack. The selected server hardware must take power density into account.
Network connectivity: The selected server hardware must have the appropriate number
of network connections, as well as the right type of network connections, in order to
support the proposed architecture. Ensure that, at a minimum, there are at least two diverse
network connections coming into each rack. For architectures requiring even
more redundancy, it might be necessary to confirm that the network connections are from
diverse telecom providers. Many data centers have that capacity available.
The selection of certain form factors or architectures will affect the selection of server
hardware. For example, if the design calls for a scale-out storage architecture (For example,
leveraging Ceph, Gluster, or a similar commercial solution), then the server hardware
selection will need to be carefully considered to match the requirements set by the
commercial solution. Ensure that the selected server hardware is configured to support
enough storage capacity (or storage expandability) to match the requirements of selected
scale-out storage solution. For example, if a centralized storage solution is required, such as
a centralized storage array from a storage vendor that has infiniBand or FDDI connections,
the server hardware will need to have appropriate network adapters installed to be
compatible with the storage array vendor's specifications.
Similarly, the network architecture will have an impact on the server hardware selection and
vice versa. For example, make sure that the server is configured with enough additional
network ports and expansion cards to support all of the networks required. There
is variability in network expansion cards, so it is important to be aware of potential impacts
or interoperability issues with other components in the architecture. This is especially true if
the architecture uses InfiniBand or another less commonly used networking protocol.

Selecting Storage Hardware

The selection of storage hardware is largely determined by the proposed storage
architecture. Factors that need to be incorporated into the storage architecture include:
Cost: Storage can be a significant portion of the overall system cost that should be
factored into the design decision. For an organization that is concerned with vendor support,
a commercial storage solution is advisable, although it is comes with a higher price tag. If
initial capital expenditure requires minimization, designing a system based on commodity

hardware would apply. The trade-off is potentially higher support costs and a greater risk of
incompatibility and interoperability issues.
Performance: Storage performance, measured by observing the latency of storage I-O
requests, is not a critical factor for a general purpose OpenStack cloud as overall systems
performance is not a design priority.
Scalability: The term "scalability" refers to how well the storage solution performs as it
expands up to its maximum designed size. A solution that continues to perform well at
maximum expansion is considered scalable. A storage solution that performs well in small
configurations but has degrading performance as it expands was not designed to be not
scalable. Scalability, along with expandability, is a major consideration in a general purpose
OpenStack cloud. It might be difficult to predict the final intended size of the
implementation because there are no established usage patterns for a general purpose
cloud. Therefore, it may become necessary to expand the initial deployment in order to
accommodate growth and user demand. The ability of the storage solution to continue to
perform well as it expands is important.
Expandability: This refers to the overall ability of the solution to grow. A storage solution
that expands to 50 PB is considered more expandable than a solution that only scales to 10
PB. This metric is related to, but different, from scalability, which is a measure of the
solution's performance as it expands. Expandability is a major architecture factor for storage
solutions with general purpose OpenStack cloud. For example, the storage architecture for a
cloud that is intended for a development platform may not have the same expandability and
scalability requirements as a cloud that is intended for a commercial product.
Storage hardware architecture is largely determined by the selected storage architecture.
The selection of storage architecture, as well as the corresponding storage hardware, is
determined by evaluating possible solutions against the critical factors, the user
requirements, technical considerations, and operational considerations. A combination of all
the factors and considerations will determine which approach will be best.
Using a scale-out storage solution with direct-attached storage (DAS) in the servers is well
suited for a general purpose OpenStack cloud. In this scenario, it is possible to populate
storage in either the compute hosts similar to a grid computing solution or into hosts
dedicated to providing block storage exclusively. When deploying storage in the compute
hosts, appropriate hardware which can support both the storage and compute services on
the same hardware will be required. This approach is referred to as a grid computing
architecture because there is a grid of modules that have both compute and storage in a
single box.

Understanding the requirements of cloud services will help determine if Ceph, Gluster, or a
similar scale-out solution should be used. It can then be further determined if a single, highly
expandable and highly vertical, scalable, centralized storage array should be included in the
design. Once the approach has been determined, the storage hardware needs to be chosen
based on this criteria. If a centralized storage array fits the requirements best, then the array
vendor will determine the hardware. For cost reasons it may be decided to build an open
source storage array using solutions such as OpenFiler, Nexenta Open Source, or BackBlaze
Open Source.
This list expands upon the potential impacts for including a particular storage architecture
(and corresponding storage hardware) into the design for a general purpose OpenStack
cloud:
Connectivity: Ensure that, if storage protocols other than Ethernet are part of the storage
solution, the appropriate hardware has been selected. Some examples include InfiniBand,
FDDI and Fibre Channel. If a centralized storage array is selected, ensure that the hypervisor
will be able to connect to that storage array for image storage.
Usage: How the particular storage architecture will be used is critical for determining the
architecture. Some of the configurations that will influence the architecture include whether
it will be used by the hypervisors for ephemeral instance storage or if OpenStack Swift will
use it for object storage. All of these usage models are affected by the selection of particular
storage architecture and the corresponding storage hardware to support that architecture.
Instance and image locations: Where instances and images will be stored will influence
the architecture. For example, instances can be stored in a number of options. OpenStack
Cinder is a good location for instances because it is persistent block storage, however, Swift
can be used if storage latency is less of a concern. The same argument applies to the
appropriate image storage location.
Server Hardware: If the solution is a scale-out storage architecture that includes DAS,
naturally that will affect the server hardware selection. This could ripple into the decisions
that affect host density, instance density, power density, OS-hypervisor, management tools
and others.
General purpose OpenStack cloud has multiple options. As a result, there is no single
decision that will apply to all implementations. The key factors that will have an influence on
selection of storage hardware for a general purpose OpenStack cloud are as follows:
Capacity: Hardware resources selected for the resource nodes should be capable of
supporting enough storage for the cloud services that will use them. It is important to clearly
define the initial requirements and ensure that the design can support adding capacity as

resources are used in the cloud, as workloads are relatively unknown. Hardware nodes
selected for object storage should be capable of supporting a large number of
inexpensive disks and should not have any reliance on RAID controller cards. Hardware
nodes selected for block storage should be capable of supporting higher speed storage
solutions and RAID controller cards to provide performance and redundancy to storage at
the hardware level. Selecting hardware RAID controllers that can automatically repair
damaged arrays will further assist with replacing and repairing degraded or destroyed
storage devices within the cloud.
Performance: Disks selected for the object storage service do not need to be fast
performing disks. It is recommended that object storage nodes take advantage of the best
cost per terabyte available for storage at the time of acquisition and avoid enterprise class
drives. In contrast, disks chosen for the block storage service should take advantage of
performance boosting features and may entail the use of SSDs or flash storage to provide
for high performing block storage pools. Storage performance of ephemeral disks used for
instances should also be taken into consideration. If compute pools are expected to have a
high utilization of ephemeral storage or requires very high performance, it would be
advantageous to deploy similar hardware solutions to block storage in order to increase the
storage performance.
Fault Tolerance: Object storage resource nodes have no requirements for hardware fault
tolerance or RAID controllers. It is not necessary to plan for fault tolerance within the object
storage hardware because the object storage service provides replication between zones as
a feature of the service. Block storage nodes, compute nodes and cloud controllers should
all have fault tolerance built in at the hardware level by making use of hardware RAID
controllers and varying levels of RAID configuration. The level of RAID chosen should be
consistent with the performance and availability requirements of the cloud.

Selecting Networking Hardware

As is the case with storage architecture, selecting a network architecture often determines
which network hardware will be used. The networking software in use is determined by the
selected networking hardware. Some design impacts are obvious, for example, selecting
networking hardware that only supports Gigabit Ethernet (GbE) will naturally have an impact
on many different areas of the overall design. Similarly, deciding to use 10 Gigabit Ethernet
(10 GbE) has a number of impacts on various areas of the overall design.
As an example, selecting Cisco networking hardware implies that the architecture will be
using Cisco networking software (IOS, NX-OS, etc.). Conversely, selecting Arista networking
hardware means the network devices will use Arista networking software (EOS). In

addition, there are more subtle design impacts that need to be considered. The selection of
certain networking hardware (and therefore the networking software) could affect the
management tools that can be used. There are exceptions to this; the rise of "open"
networking software that supports a range of networking hardware means that there are
instances where the relationship between networking hardware and networking software
are not as tightly defined. An example of this type of software is Cumulus Linux, which is
capable of running on a number of switch vendor’s hardware solutions.
Some of the key considerations that should be included in the selection of networking
hardware include:
Port count: The design will require networking hardware that has the requisite port count.
Port density: The network design will be affected by the physical space that is required to
provide the requisite port count. A switch that can provide 48 10 GbE ports in 1U has a much
higher port density than a switch that provides 24 10 GbE ports in 2U. A higher port density
is preferred, as it leaves more rack space for compute or storage components that may be
required by the design. This can also lead into concerns about fault domains and power
density that should be considered. Higher density switches are more expensive and should
also be considered, as it is important not to over design the network if it is not required.
Port speed: The networking hardware must support the proposed network speed, for
example: 1 GbE, 10 GbE, or 40 GbE (or even 100 GbE).
Redundancy: The level of network hardware redundancy required is influenced by the
user requirements for high availability and cost considerations. Network redundancy can be
achieved by adding redundant power supplies or paired switches. If this is a requirement, the
hardware will need to support this configuration. User requirements will determine if a
completely redundant network infrastructure is required.
Power requirements: Make sure that the physical data center provides the necessary
power for the selected network hardware. This is not an issue for top of rack (ToR) switches,
but may be an issue for spine switches in a leaf and spine fabric, or end of row (EoR)
switches.
There is no single best practice architecture for the networking hardware supporting a
general purpose OpenStack cloud that will apply to all implementations. Some of the key
factors that will have a strong influence on selection of networking hardware include:
Connectivity: All nodes within an OpenStack cloud require some form of network
connectivity. In some cases, nodes require access to more than one network segment. The
design must encompass sufficient network capacity and bandwidth to ensure that all
communications within the cloud, both north-south and east-west traffic have sufficient

resources available.
Scalability: The chosen network design should encompass a physical and logical network
design that can be easily expanded upon. Network hardware should offer the appropriate
types of interfaces and speeds that are required by the hardware nodes.
Availability: To ensure that access to nodes within the cloud is not interrupted, it is
recommended that the network architecture identify any single points of failure and provide
some level of redundancy or fault tolerance. With regard to the network infrastructure itself,
this often involves use of networking protocols such as LACP, VRRP or others to achieve a
highly available network connection. In addition, it is important to consider the networking
implications on API availability. In order to ensure that the APIs, and potentially other
services in the cloud are highly available, it is recommended to design load balancing
solutions within the network architecture to accommodate for these requirements.

Software Selection

Software selection for a general purpose OpenStack architecture design needs to include
these three areas:
Operating system (OS) and hypervisor
OpenStack components
Supplemental software

OS and Hypervisor

The selection of OS and hypervisor has a tremendous impact on the overall design. Selecting
a particular operating system and hypervisor can also directly affect server hardware
selection. It is recommended to make sure the storage hardware selection and topology
support the selected operating system and hypervisor combination. Finally, it is important
to ensure that the networking hardware selection and topology will work with the chosen
operating system and hypervisor combination. For example, if the design uses Link
Aggregation Control Protocol (LACP), the OS and hypervisor both need to support it.
Some areas that could be impacted by the selection of OS and hypervisor include:
Cost: Selecting a commercially supported hypervisor, such as Microsoft Hyper-V, will
result in a different cost model rather than community-supported open source hypervisors
including KVM, Kinstance or Xen. When comparing open source OS solutions, choosing
Ubuntu over Red Hat (or vice versa) will have an impact on cost due to support contracts. On
the other hand, business or application requirements may dictate a specific or commercially
supported hypervisor.
Supportability: Depending on the selected hypervisor, the staff should have the

appropriate training and knowledge to support the selected OS and hypervisor combination.
If they do not, training will need to be provided which could have a cost impact on the
design.
Management tools: The management tools used for Ubuntu and Kinstance differ
from the management tools for VMware vSphere. Although both OS and hypervisor
combinations are supported by OpenStack, there will be very different impacts to the rest of
the design as a result of the selection of one combination versus the other.
Scale and performance: Ensure that selected OS and hypervisor combinations meet the
appropriate scale and performance requirements. The chosen architecture will need to meet
the targeted instance-host ratios with the selected OS-hypervisor combinations.
Security: Ensure that the design can accommodate the regular periodic installation of
application security patches while maintaining the required workloads. The frequency of
security patches for the proposed OS - hypervisor combination will have an impact on
performance and the patch installation process could affect maintenance windows.
Supported features: Determine which features of OpenStack are required. This will often
determine the selection of the OS-hypervisor combination. Certain features are only
available with specific OSs or hypervisors. For example, if certain features are not available,
the design might need to be modified to meet the user requirements.
Interoperability: Consideration should be given to the ability of the selected OShypervisor combination to interoperate or co-exist with other OS-hypervisors as well as
other software solutions in the overall design (if required). Operational troubleshooting tools
for one OS-hypervisor combination may differ from the tools used for another OShypervisor combination and, as a result, the design will need to address if the two sets of
tools need to interoperate.

OpenStack Components

The selection of which OpenStack components are included has a significant impact on the
overall design. While there are certain components that will always be present, (Nova and
Glance, for example) there are other services that may not be required. As an example, a
certain design might not need OpenStack Heat. Omitting Heat would not have a significant
impact on the overall design of a cloud; however, if the architecture uses a replacement
for OpenStack Swift for its storage component, it could potentially have significant impacts
on the rest of the design.
The exclusion of certain OpenStack components might also limit or constrain the
functionality of other components. If the architecture includes Heat but excludes
Ceilometer, then the design will not be able to take advantage of Heat's auto scaling

functionality (which relies on information from Ceilometer). It is important to research the
component interdependencies in conjunction with the technical requirements before
deciding what components need to be included and what components can be dropped from
the final architecture.

Supplemental Components

While OpenStack is a fairly complete collection of software projects for building a platform
for cloud services, there are invariably additional pieces of software that need to be
considered in any given OpenStack design.

Networking Software

OpenStack Neutron provides a wide variety of networking services for instances. There are
many additional networking software packages that might be useful to manage the
OpenStack components themselves. Some examples include software to provide load
balancing, network redundancy protocols, and routing daemons. Some of these software
packages are described in more detail in the OpenStack HA Guide (refer to Chapter 8 of the
OpenStack High Availability Guide).
For a general purpose OpenStack cloud, the OpenStack infrastructure components will need
to be highly available. If the design does not include hardware load balancing, networking
software packages like HAProxy will need to be included.

Management Software

The selected supplemental software solution impacts and affects the overall OpenStack
cloud design. This includes software for providing clustering, logging, monitoring and
alerting.
Inclusion of clustering Software, such as Corosync or Pacemaker, is determined primarily by
the availability requirements. Therefore, the impact of including (or not including) these
software packages is primarily determined by the availability of the cloud infrastructure and
the complexity of supporting the configuration after it is deployed. The OpenStack High
Availability Guide provides more details on the installation and configuration of Corosync
and Pacemaker, should these packages need to be included in the design.
Requirements for logging, monitoring, and alerting are determined by operational
considerations. Each of these sub-categories includes a number of various options. For
example, in the logging sub-category one might consider Logstash, Splunk, instanceware
Log Insight, or some other log aggregation-consolidation tool. Logs should be stored in a
centralized location to make it easier to perform analytics against the data. Log data

analytics engines can also provide automation and issue notification by providing a
mechanism to both alert and automatically attempt to remediate some of the more
commonly known issues.
If any of these software packages are required, then the design must account for the
additional resource consumption (CPU, RAM, storage, and network bandwidth for a log
aggregation solution, for example). Some other potential design impacts include:
OS - Hypervisor combination: Ensure that the selected logging, monitoring, or alerting
tools support the proposed OS-hypervisor combination.
Network hardware: The network hardware selection needs to be supported by
the logging, monitoring, and alerting software.

Database Software

A large majority of the OpenStack components require access to back-end database
services to store state and configuration information. Selection of an appropriate back-end
database that will satisfy the availability and fault tolerance requirements of the OpenStack
services is required. OpenStack services supports connecting to any database that is
supported by the sqlalchemy python drivers, however, most common database
deployments make use of mySQL or variations of it. It is recommended that the database
which provides back-end service within a general purpose cloud be made highly available
when using an available technology which can accomplish that goal. Some of the more
common software solutions used include Galera, MariaDB and mySQL with multi-master
replication.

Addressing Performance-Sensitive Workloads

Although one of the key defining factors for a general purpose OpenStack cloud is that
performance is not a determining factor, there may still be some performance-sensitive
workloads deployed on the general purpose OpenStack cloud. For design guidance on
performance-sensitive workloads, it is recommended to refer to the focused scenarios later
in this guide. The resource-focused guides can be used as a supplement to this guide to help
with decisions regarding performance-sensitive workloads.

Compute-Focused Workloads

In an OpenStack cloud that is compute-focused, there are some design choices that can
help accommodate those workloads. Compute-focused workloads are generally those that
would place a higher demand on CPU and memory resources with lower priority given
to storage and network performance, other than what is required to support the intended
compute workloads. For guidance on designing for this type of cloud, please refer to the

section on Compute Focused clouds.

Network-Focused Workloads

In a network-focused OpenStack cloud some design choices can improve the performance
of these types of workloads. Network-focused workloads have extreme demands on
network bandwidth and services that require specialized consideration and planning. For
guidance on designing for this type of cloud, please refer to the section on NetworkFocused clouds.

Storage-Focused Workloads

Storage focused OpenStack clouds need to be designed to accommodate workloads that
have extreme demands on either object or block storage services that require specialized
consideration and planning. For guidance on designing for this type of cloud, please refer to
the section on Storage-Focused clouds.

Prescriptive Example

An online classified advertising company wants to run web applications consisting of
Tomcat, Nginx and MariaDB in a private cloud. In order to meet policy requirements, the
cloud infrastructure will run in their own datacenter. They have predictable load
requirements but require an element of scaling to cope with nightly increases in demand.
Their current environment is not flexible enough to align with their goal of running an open
source API driven environment. Their current environment consists of the following:
Between 120 and 140 installations of Nginx and Tomcat, each with 2 vCPUs and 4 GB of
RAM
A three-node MariaDB and Galera cluster, each with 4 vCPUs and 8 GB RAM
The company runs hardware load balancers and multiple web applications serving the
sites. The company orchestrates their environment using a combination of scripts and
Puppet. The websites generate a large amount of log data each day that needs to be
archived.
The solution would consist of the following OpenStack components:
A firewall, switches and load balancers on the public facing network connections.
OpenStack Controller services running Image, Identity, Networking and supporting
services such as MariaDB and RabbitMQ. The controllers will run in a highly available
configuration on at least three controller nodes.
OpenStack Compute nodes running the KVM hypervisor.
OpenStack Block Storage for use by compute instances that require persistent storage
such as databases for dynamic sites.
OpenStack Object Storage for serving static objects such as images.

Running up to 140 web instances and the small number of MariaDB instances requires 292
vCPUs available, as well as 584 GB RAM. On a typical 1U server using dual-socket hex-core
Intel CPUs with Hyperthreading, and assuming 2:1 CPU overcommit ratio, this would require
8 OpenStack Compute nodes.

The web application instances run from local storage on each of the OpenStack Compute
nodes. The web application instances are stateless, meaning that any of the instances can
fail and the application will continue to function.
MariaDB server instances store their data on shared enterprise storage, such as NetApp or
Solidfire devices. If a MariaDB instance fails, storage would be expected to be re-attached to
another instance and rejoined to the Galera cluster.
Logs from the web application servers are shipped to OpenStack Object Storage for later
processing and archiving.
In this scenario, additional capabilities can be realized by moving static web content to be
served from OpenStack Object Storage containers, and backing the OpenStack Image
Service with OpenStack Object Storage. Note that an increase in OpenStack Object Storage
means that network bandwidth needs to be taken in to consideration. It is best to run
OpenStack Object Storage with network connections offering 10 GbE or better connectivity.
There is also a potential to leverage the Orchestration and Telemetry OpenStack modules to
provide an auto-scaling, orchestrated web application environment. Defining the web
applications in Heat Orchestration Templates (HOT) would negate the reliance on the
scripted Puppet solution currently employed.
OpenStack Networking can be used to control hardware load balancers through the use of
plugins and the Networking API. This would allow a user to control hardware load balance
pools and instances as members in these pools, but their use in production environments
must be carefully weighed against current stability.

COMPUTE FOCUSED

Introduction

A compute-focused cloud is a specialized subset of the general purpose OpenStack cloud
architecture. Unlike the general purpose OpenStack architecture, which is built to host a
wide variety of workloads and applications and does not heavily tax any particular
computing aspect, a compute-focused cloud is built and designed specifically to
support compute intensive workloads. As such, the design must be specifically tailored
to support hosting compute intensive workloads. Compute intensive workloads may be
CPU intensive, RAM intensive, or both. However, they are not typically storage intensive or
network intensive. Compute-focused workloads may include the following use cases:
High performance computing (HPC)
Big data analytics using Hadoop or other distributed data stores
Continuous integration/continuous deployment (CI/CD)
Platform-as-a-Service (PaaS)
Signal processing for Network Function Virtualization (NFV)
Based on the use case requirements, such clouds might need to provide additional services
such as a virtual machine disk library, file or object storage, firewalls, load balancers, IP
addresses, and network connectivity in the form of overlays or virtual Local Area Networks
(VLANs). A compute-focused OpenStack cloud will not typically use raw block storage
services since the applications hosted on a compute-focused OpenStack cloud generally do
not need persistent block storage.

User Requirements

Compute intensive workloads are defined by their high utilization of CPU, RAM, or both. User
requirements will determine if a cloud must be built to accommodate anticipated
performance demands.

Business Considerations

Cost: Cost is not generally a primary concern for a compute-focused cloud, however
some organizations might be concerned with cost avoidance. Repurposing existing
resources to tackle compute-intensive tasks instead of needing to acquire additional
resources may offer cost reduction opportunities.
Time to Market: Compute-focused clouds can be used to deliver products more quickly,
for example, speeding up a company's software development lifecycle (SDLC) for building
products and applications.
Revenue Opportunity: Companies that are interested in building services or products that
rely on the power of the compute resources will benefit from a compute-focused
cloud. Examples include the analysis of large data sets (via Hadoop or Cassandra)
or completing computational intensive tasks such as rendering, scientific computation, or
simulations.

Legal Requirements

Many jurisdictions have legislative and regulatory requirements governing the storage and
management of data in cloud environments. Common areas of regulation include:
Data retention policies ensuring storage of persistent data and records management to
meet data archival requirements.
Data ownership policies governing the possession and responsibility for data.
Data sovereignty policies governing the storage of data in foreign countries or otherwise
separate jurisdictions.
Data compliance - certain types of information needs to reside in certain locations due to
regular issues - and more important cannot reside in other locations for the same reason.
Examples of such legal frameworks include the data protection framework of the European
Union (http://ec.europa.eu/justice/data-protection/ ) and the requirements of the Financial
Industry Regulatory Authority (http://www.finra.org/Industry/Regulation/FINRARules/ ) in
the United States. Consult a local regulatory body for more information.

Technical Considerations

The following are some technical requirements that need to be incorporated into the
architecture design.
Performance: If a primary technical concern is for the environment to deliver high
performance capability, then a compute-focused design is an obvious choice because it is
specifically designed to host compute-intensive workloads.
Workload persistence: Workloads can be either short-lived or long running. Short-lived
workloads might include continuous integration and continuous deployment (CI-CD) jobs,
where large numbers of compute instances are created simultaneously to perform a set of
compute-intensive tasks. The results or artifacts are then copied from the instance into
long-term storage before the instance is destroyed. Long-running workloads, like a Hadoop
or high-performance computing (HPC) cluster, typically ingest large data sets, perform the
computational work on those data sets, then push the results into long term storage. Unlike
short-lived workloads, when the computational work is completed, they will remain idle
until the next job is pushed to them. Long-running workloads are often larger and more
complex, so the effort of building them is mitigated by keeping them active between jobs.
Another example of long running workloads is legacy applications that typically are
persistent over time.
Storage: Workloads targeted for a compute-focused OpenStack cloud generally do not
require any persistent block storage (although some usages of Hadoop with HDFS may
dictate the use of persistent block storage). A shared filesystem or object store will maintain
the initial data set(s) and serve as the destination for saving the computational results. By
avoiding the input-output (IO) overhead, workload performance is significantly
enhanced. Depending on the size of the data set(s), it might be necessary to scale the object
store or shared file system to match the storage demand.
User Interface: Like any other cloud architecture, a compute-focused OpenStack
cloud requires an on-demand and self-service user interface. End users must be able to
provision computing power, storage, networks and software simply and flexibly. This
includes scaling the infrastructure up to a substantial level without disrupting host
operations.
Security: Security is going to be highly dependent on the business requirements. For
example, a computationally intense drug discovery application will obviously have much
higher security requirements than a cloud that is designed for processing market data for a
retailer. As a general start, the security recommendations and guidelines provided in the
OpenStack Security Guide are applicable.

Operational Considerations

The compute intensive cloud from the operational perspective is similar to the requirements

for the general-purpose cloud. More details on operational requirements can be found in
the general-purpose design section.

Technical Considerations

In a compute-focused OpenStack cloud, the type of instance workloads being provisioned
heavily influences technical decision making. For example, specific use cases that demand
multiple short running jobs present different requirements than those that specify longrunning jobs, even though both situations are considered "compute focused."

Capacity

Public and private clouds require deterministic capacity planning to support elastic growth in
order to meet user SLA expectations. Deterministic capacity planning is the path to
predicting the effort and expense of making a given process consistently performant. This
process is important because, when a service becomes a critical part of a user's
infrastructure, the user's fate becomes wedded to the SLAs of the cloud itself. In cloud
computing, a service’s performance will not be measured by its average speed but rather by
the consistency of its speed.
There are two aspects of capacity planning to consider: planning the initial deployment
footprint, and planning expansion of it to stay ahead of the demands of cloud users.

Initial Planning

Planning the initial footprint for an OpenStack deployment is typically done based on
existing infrastructure workloads and estimates based on expected uptake.
The starting point is the core count of the cloud. By applying relevant ratios, the user can
gather information about:
The number of instances expected to be available concurrently: (overcommit fraction ×
cores) / virtual cores per instance
How much storage is required: flavor disk size × number of instances
These ratios can be used to determine the amount of additional infrastructure needed to
support the cloud. For example, consider a situation in which you require 1600 instances,
each with 2 vCPU and 50 GB of storage. Assuming the default overcommit rate of 16:1,
working out the math provides an equation of:
1600 = (16 x (number of physical cores)) / 2
storage required = 50 GB x 1600
On the surface, the equations reveal the need for 200 physical cores and 80 TB of storage
for /var/lib/nova/instances/. However, it is also important to look at patterns of usage to

estimate the load that the API services, database servers, and queue servers are likely to
encounter.
Consider, for example, the differences between a cloud that supports a managed webhosting platform with one running integration tests for a development project that creates
one instance per code commit. In the former, the heavy work of creating an
instance happens only every few months, whereas the latter puts constant heavy load on
the cloud controller. The average instance lifetime must be considered, as a larger number
generally means less load on the cloud controller.
Aside from the creation and termination of instances, the impact of users must be
considered when accessing the service, particularly on nova-api and its associated database.
Listing instances garners a great deal of information and, given the frequency with which
users run this operation, a cloud with a large number of users can increase the load
significantly. This can even occur unintentionally. For example, the OpenStack Dashboard
instances tab refreshes the list of instances every 30 seconds, so leaving it open in a browser
window can cause unexpected load.
Consideration of these factors can help determine how many cloud controller cores
are required. A server with 8 CPU cores and 8 GB of RAM server would be sufficient for up to
a rack of compute nodes, given the above caveats.
Key hardware specifications are also crucial to the performance of user instances. Be sure to
consider budget and performance needs, including storage performance (spindles/core),
memory availability (RAM/core), network bandwidth (Gbps/core), and overall CPU
performance (CPU/core).
The cloud resource calculator is a useful tool in examining the impacts of different hardware
and instance load outs. It is available at:
https://github.com/noslzzp/cloud-resource-calculator/blob/master/cloudresource-calculator.ods

Expansion Planning

A key challenge faced when planning the expansion of cloud compute services is the elastic
nature of cloud infrastructure demands. Previously, new users or customers would be
forced to plan for and request the infrastructure they required ahead of time, allowing time
for reactive procurement processes. Cloud computing users have come to expect the agility
provided by having instant access to new resources as they are required. Consequently, this

means planning should be delivered for typical usage, but also more importantly, for sudden
bursts in usage.
Planning for expansion can be a delicate balancing act. Planning too conservatively can lead
to unexpected oversubscription of the cloud and dissatisfied users. Planning for cloud
expansion too aggressively can lead to unexpected underutilization of the cloud and funds
spent on operating infrastructure that is not being used efficiently.
The key is to carefully monitor the spikes and valleys in cloud usage over time. The intent is
to measure the consistency with which services can be delivered, not the average speed or
capacity of the cloud. Using this information to model performance results in
capacity enables users to more accurately determine the current and future capacity of the
cloud.

CPU and RAM

(Adapted from: http://docs.openstack.org/openstackops/content/compute_nodes.html#cpu_choice)
In current generations, CPUs have up to 12 cores. If an Intel CPU supports Hyper-Threading,
those 12 cores are doubled to 24 cores. If a server is purchased that supports multiple CPUs,
the number of cores is further multiplied. Hyper-Threading is Intel's proprietary
simultaneous multi-threading implementation, used to improve parallelization on their
CPUs. Consider enabling Hyper-Threading to improve the performance of multithreaded
applications.
Whether the user should enable Hyper-Threading on a CPU depends upon the use case. For
example, disabling Hyper-Threading can be beneficial in intense computing environments.
Performance testing conducted by running local workloads with both Hyper-Threading on
and off can help determine what is more appropriate in any particular case.
If the Libvirt/KVM Hypervisor driver are the intended use cases, then the CPUs used in the
compute nodes must support virtualization by way of the VT-x extensions for Intel chips and
AMD-v extensions for AMD chips to provide full performance.
OpenStack enables the user to overcommit CPU and RAM on compute nodes. This allows an
increase in the number of instances running on the cloud at the cost of reducing the
performance of the instances. OpenStack Compute uses the following ratios by default:
CPU allocation ratio: 16:1
RAM allocation ratio: 1.5:1

The default CPU allocation ratio of 16:1 means that the scheduler allocates up to 16 virtual
cores per physical core. For example, if a physical node has 12 cores, the scheduler sees 192
available virtual cores. With typical flavor definitions of 4 virtual cores per instance, this ratio
would provide 48 instances on a physical node.
Similarly, the default RAM allocation ratio of 1.5:1 means that the scheduler allocates
instances to a physical node as long as the total amount of RAM associated with the
instances is less than 1.5 times the amount of RAM available on the physical node.
For example, if a physical node has 48 GB of RAM, the scheduler allocates instances to that
node until the sum of the RAM associated with the instances reaches 72 GB (such as nine
instances, in the case where each instance has 8 GB of RAM).
The appropriate CPU and RAM allocation ratio must be selected based on particular use
cases.

Additional Hardware

Certain use cases may benefit from exposure to additional devices on the compute node.
Examples might include:
High performance computing jobs that benefit from the availability of graphics
processing units (GPUs) for general-purpose computing.
Cryptographic routines that benefit from the availability of hardware random number
generators to avoid entropy starvation.
Database management systems that benefit from the availability of SSDs for
ephemeral storage to maximize read/write time when it is required.
Host aggregates are used to group hosts that share similar characteristics, which can include
hardware similarities. The addition of specialized hardware to a cloud deployment is likely to
add to the cost of each node, so careful consideration must be given to whether all compute
nodes, or just a subset which is targetable using flavors, need the additional customization
to support the desired workloads.

Utilization

Infrastructure-as-a-Service offerings, including OpenStack, use flavors to provide
standardized views of virtual machine resource requirements that simplify the problem of
scheduling instances while making the best use of the available physical resources.
In order to facilitate packing of virtual machines onto physical hosts, the default selection of

flavors are constructed so that the second largest flavor is half the size of the largest flavor
in every dimension. It has half the vCPUs, half the vRAM, and half the ephemeral disk space.
The next largest flavor is half that size again. As a result, packing a server for general
purpose computing might look conceptually something like Figure X:

On the other hand, a CPU optimized packed server might look like the following figure:

These default flavors are well suited to typical load outs for commodity server hardware. To
maximize utilization, however, it may be necessary to customize the flavors or create new
ones, to better align instance sizes to the available hardware.
Workload characteristics may also influence hardware choices and flavor configuration,
particularly where they present different ratios of CPU versus RAM versus HDD
requirements.
For more information on Flavors refer to: http://docs.openstack.org/openstackops/content/flavors.html

Performance

The infrastructure of a cloud should not be shared, so that it is possible for the workloads to
consume as many resources as are made available, and accommodations should be made
to provide large scale workloads.

The duration of batch processing differs depending on individual workloads that are
launched. Time limits range from seconds, minutes to hours, and as a result it is considered
difficult to predict when resources will be used, for how long, and even which resources will
be used.

Security

The security considerations needed for this scenario are similar to those of the other
scenarios discussed in this book.
A security domain comprises of users, applications, servers or networks that share common
trust requirements and expectations within a system. Typically they have the same
authentication and authorization requirements and users.
These security domains are:
1. Public
2. Guest
3. Management
4. Data
These security domains can be mapped individually to the installation, or they can also be
combined. For example, some deployment topologies combine both guest and data
domains onto one physical network, whereas in other cases these networks are physically
separated. In each case, the cloud operator should be aware of the appropriate security
concerns. Security domains should be mapped out against specific OpenStack deployment
topology. The domains and their trust requirements depend upon whether the cloud
instance is public, private, or hybrid.
The public security domain is an entirely untrusted area of the cloud infrastructure. It can
refer to the Internet as a whole or simply to networks over which the user has no
authority. This domain should always be considered untrusted.
Typically used for compute instance-to-instance traffic, the guest security domain handles
compute data generated by instances on the cloud; not services that support the operation
of the cloud, for example API calls. Public cloud providers and private cloud providers who
do not have stringent controls on instance use or who allow unrestricted internet access to
instances should consider this domain to be untrusted. Private cloud providers may want to

consider this network as internal and therefore trusted only if they have controls in place to
assert that they trust instances and all their tenants.
The management security domain is where services interact. Sometimes referred to as the
"control plane", the networks in this domain transport confidential data such as
configuration parameters, usernames, and passwords. In most deployments this domain is
considered trusted.
The data security domain is concerned primarily with information pertaining to the storage
services within OpenStack. Much of the data that crosses this network has high integrity and
confidentiality requirements and depending on the type of deployment there may also be
strong availability requirements. The trust level of this network is heavily dependent on
deployment decisions and as such we do not assign this any default level of trust.
When deploying OpenStack in an enterprise as a private cloud it is assumed to be behind a
firewall and within the trusted network alongside existing systems. Users of the cloud are
typically employees or trusted individuals that are bound by the security requirements set
forth by the company. This tends to push most of the security domains towards a more
trusted model. However, when deploying OpenStack in a public-facing role, no assumptions
can be made and the attack vectors significantly increase. For example, the API
endpoints and the software behind it will be vulnerable to potentially hostile entities
wanting to gain unauthorized access or prevent access to services. This can result in loss of
reputation and must be protected against through auditing and appropriate filtering.
Consideration must be taken when managing the users of the system, whether it is
the operation of public or private clouds. The identity service allows for LDAP to be part of
the authentication process, and includes such systems as an OpenStack deployment
that may ease user management if integrated into existing systems.
It is strongly recommended that the API services are placed behind hardware that performs
SSL termination. API services transmit usernames, passwords, and generated
tokens between client machines and API endpoints and therefore must be secured.
More information on OpenStack Security can be found at
http://docs.openstack.org/security-guide/

OpenStack Components

Due to the nature of the workloads that will be used in this scenario, a number
of components will be highly beneficial in a Compute-focused cloud. This includes the

typical OpenStack components:
OpenStack Compute (Nova)
OpenStack Image Service (Glance)
OpenStack Identity Service (Keystone)
Also consider several specialized components:
OpenStack Orchestration Engine (Heat)
It is safe to assume that, given the nature of the applications involved in this scenario,
these will be heavily automated deployments. Making use of Heat will be highly
beneficial in this case. Deploying a batch of instances and running an automated set of
tests can be scripted, however it makes sense to use the OpenStack Orchestration
Engine (Heat) to handle all these actions.
OpenStack Telemetry (Ceilometer)
OpenStack Telemetry and the alarms it generates are required to support autoscaling
of instances using OpenStack Orchestration. Users that are not using OpenStack
Orchestration do not need to deploy OpenStack Telemetry and may choose to use
other external solutions to fulfill their metering and monitoring requirements.
See also: http://docs.openstack.org/openstackops/content/logging_monitoring.html
OpenStack Block Storage (Cinder)
Due to the burstable nature of the workloads and the applications and instances
that will be used for batch processing, this cloud will utilize mainly memory or CPU, so
the need for add-on storage to each instance is not a likely requirement. This does not
mean the OpenStack Block Storage service (Cinder) will not be used in the
infrastructure, but typically it will not be used as a central component.
Networking
When choosing a networking platform, ensure that it either works with all desired
hypervisor and container technologies and their OpenStack drivers, or includes an
implementation of an ML2 mechanism driver. Networking platforms that provide ML2
mechanisms drivers can be mixed.

Operational Considerations

Operationally, there are a number of considerations that affect the design of computefocused OpenStack clouds. Some examples might include enforcing strict API availability
requirements, understanding and dealing with failure scenarios, or managing host
maintenance schedules.

Service-Level Agreements

Service-level agreements (SLAs) are a contractual obligation that gives assurances around
availability of a provided service. As such, factoring in promises of availability implies a
certain level of redundancy and resiliency when designing an OpenStack cloud.
Guarantees for API availability imply multiple infrastructure services combined with
appropriately high available load balancers.
Network uptime guarantees will affect the switch design and might
require redundant switching and power.
Network security policy requirements need to be factored in to deployments.
Knowing when and where to implement redundancy and high availability (HA) is directly
affected by terms contained in any associated SLA, if one is present.

Support and Maintainability

OpenStack cloud management requires operations staff to be able to understand and
comprehend design architecture content on some level. The level of skills and the level of
separation of the operations and engineering staff is dependent on the size and purpose of
the installation. A large cloud service provider or a telecom provider is more inclined to be
managed by specially trained dedicated operations organization. A smaller implementation
is more inclined to rely on a smaller support staff that might need to take on the combined
engineering, design and operations functions.
Maintaining OpenStack installations require a variety of technical skills. Some of these
skills may include the ability to debug Python log output to a basic level as well as an
understanding of networking concepts.
Consider incorporating features into the architecture and design that reduce the
operational burden. Some examples include automating some of the operations functions,
or alternatively exploring the possibility of using a third party management company with
special expertise in managing OpenStack deployments.

Monitoring

Like any other infrastructure deployment, OpenStack clouds need an appropriate
monitoring platform to ensure errors are caught and managed appropriately. Consider
leveraging any existing monitoring system to see if it will be able to effectively monitor an
OpenStack environment. While there are many aspects that need to be monitored, specific
metrics that are critically important to capture include image disk utilization, or response
time to the Nova API.

Expected and unexpected server downtime

At some point, servers will fail. The SLAs in place affect how the design has to
address recovery time. Recovery of a failed host may mean restoring instances from a
snapshot, or respawning that instance on another available host, which then has
consequences on the overall application design running on the OpenStack cloud.
It might be acceptable to design a compute-focused cloud without the ability to migrate
instances from one host to another, because the expectation is that the application
developer must handle failure within the application itself. Conversely, a compute-focused
cloud might be provisioned to provide extra resilience as a requirement of that business. In
this scenario, it is expected that extra supporting services are also deployed, such as shared
storage attached to hosts to aid in recovery and resiliency of services in order to meet strict
SLAs.

Capacity Planning

Adding extra capacity to an OpenStack cloud is an easy horizontally scaling process, as
consistently configured nodes automatically attach to an OpenStack cloud. Be mindful,
however, of any additional work to place the nodes into appropriate Availability Zones and
Host Aggregates if necessary. The same (or very similar) CPUs are recommended when
adding extra nodes to the environment because it reduces the chance to break any livemigration features if they are present. Scaling out hypervisor hosts also has a direct effect
on network and other data center resources, so factor in this increase when reaching rack
capacity or when extra network switches are required.
Compute hosts can also have internal components changed to account for increases in
demand, a process also known as vertical scaling. Swapping a CPU for one with more
cores, or increasing the memory in a server, can help add extra needed capacity depending
on whether the running applications are more CPU intensive or memory based (as would be
expected in a compute-focused OpenStack cloud).

Another option is to assess the average workloads and increase the number of instances
that can run within the compute environment by adjusting the overcommit ratio. While only
appropriate in some environments, it's important to remember that changing the CPU
overcommit ratio can have a detrimental effect and cause a potential increase in noisy
neighbor. The added risk of increasing the overcommit ratio is more instances will fail when
a compute host fails. In a compute-focused OpenStack design architecture, increasing the
CPU overcommit ratio increases the potential for noisy neighbor issues and is not
recommended.

Architecture

Hardware Selection

The hardware selection covers three areas:
Compute
Network
Storage
In a compute-focused OpenStack cloud the hardware selection must reflect the workloads
being compute intensive. Compute-focused is defined as having extreme demands on
processor and memory resources. The hardware selection for a compute-focused
OpenStack architecture design must reflect this preference for compute-intensive
workloads, as these workloads are not storage intensive, nor are they consistently network
intensive. The network and storage may be heavily utilized while loading a data set into the
computational cluster, but they are not otherwise intensive.

Compute (Server) Hardware Selection

Compute (server) hardware must be evaluated against four opposing dimensions:
Server density: A measure of how many servers can fit into a given measure of physical
space, such as a rack unit [U].
Resource capacity: The number of CPU cores, how much RAM, or how much storage a
given server will deliver.
Expandability: The number of additional resources that can be added to a server before it
has reached its limit.
Cost: The relative purchase price of the hardware weighted against the level of design
effort needed to build the system.
The dimensions need to be weighted against each other to determine the best design for the
desired purpose. For example, increasing server density means sacrificing resource capacity
or expandability. Increasing resource capacity and expandability can increase cost but
decreases server density. Decreasing cost can mean decreasing supportability, server
density, resource capacity, and expandability.
Selection of hardware for a compute-focused cloud should have an emphasis on server
hardware that can offer more CPU sockets, more CPU cores, and more RAM; network
connectivity and storage capacity are less critical. The hardware will need to be configured
to provide enough network connectivity and storage capacity to meet minimum user
requirements, but they are not the primary consideration.

Some server hardware form factors are better suited than others, as CPU and RAM capacity
have the highest priority.
Most blade servers can support dual-socket multi-core CPUs. To avoid the limit means
selecting "full width" or "full height" blades, which consequently loses server density. As an
example, using high density blade servers including HP BladeSystem and Dell PowerEdge
M1000e) which support up to 16 servers in only 10 rack units using halfheight blades, suddenly decreases the density by 50% by selecting full-height blades
resulting in only 8 servers per 10 rack units.
1U rack-mounted servers (servers that occupy only a single rack unit) may be able to
offer greater server density than a blade server solution. It is possible to place 40 servers in a
rack, providing space for the top of rack [ToR] switches, versus 32 "full width" or "full height"
blade servers in a rack), but often are limited to dual-socket, multi-core CPU configurations.
Note that, as of the Icehouse release, neither HP, IBM, nor Dell offered 1U rack servers with
more than 2 CPU sockets. To obtain greater than dual-socket support in a 1U rack-mount
form factor, customers need to buy their systems from Original Design Manufacturers
(ODMs) or second-tier manufacturers. This may cause issues for organizations that have
preferred vendor policies or concerns with support and hardware warranties of non-tier 1
vendors.
2U rack-mounted servers provide quad-socket, multi-core CPU support, but with a
corresponding decrease in server density (half the density offered by 1U rack-mounted
servers).
Larger rack-mounted servers, such as 4U servers, often provide even greater CPU
capacity, commonly supporting four or even eight CPU sockets. These servers have greater
expandability, but such servers have much lower server density and usually greater
hardware cost.
"Sled servers" (rack-mounted servers that support multiple independent servers in a
single 2U or 3U enclosure) deliver increased density as compared to typical 1U or 2U rackmounted servers. For example, many sled servers offer four independent dual-socket
nodes in 2U for a total of 8 CPU sockets in 2U. However, the dual-socket limitation on
individual nodes may not be sufficient to offset their additional cost and configuration
complexity.
The following facts will strongly influence server hardware selection for a compute-focused
OpenStack design architecture:
Instance density: In this architecture instance density is considered lower; therefore CPU
and RAM oversubscription ratios are also lower. More hosts will be required to support the
anticipated scale due to instance density being lower, especially if the design uses dual-

socket hardware designs.
Host density: Another option to address the higher host count that might be needed with
dual socket designs is to use a quad socket platform. Taking this approach will decrease
host density, which will have the access access of increasing rack count. This configuration
may affect the network requirements, the number of power connections, and
possibly impact the cooling requirements.
Power and cooling density: The power and cooling density requirements might be lower
than with blade, sled, or 1U server designs because of lower host density (by using 2U, 3U or
even 4U server designs). For data centers with older infrastructure, this may be a desirable
feature.
Compute-focused OpenStack design architecture server hardware selection results in a
"scale up" versus "scale out" decision. Selecting a better solution, smaller number of larger
hosts, or a larger number of smaller hosts depends on a combination of factors: cost, power,
cooling, physical rack and floor space, support-warranty, and manageability.

Storage Hardware Selection

For a compute-focused OpenStack design architecture, the selection of storage hardware is
not critical as it is not primary criteria, however it is still important. There are a number of
different factors that a cloud architect must consider:
Cost: The overall cost of the solution will play a major role in what storage architecture
(and resulting storage hardware) is selected.
Performance: The performance of the solution is also a big role and can be measured by
observing the latency of storage I-O requests. In a compute-focused OpenStack cloud,
storage latency can be a major consideration. In some compute-intensive workloads,
minimizing the delays that the CPU experiences while fetching data from the storage can
have a significant impact on the overall performance of the application.
Scalability: This section will refer to the term "scalability" to refer to how well the storage
solution performs as it is expanded up to its maximum size. A storage solution that performs
well in small configurations but has degrading performance as it expands would not
be considered scalable. On the other hand, a solution that continues to perform well at
maximum expansion would be considered scalable.
Expandability: Expandability refers to the overall ability of the solution to grow. A storage
solution that expands to 50 PB is considered more expandable than a solution that only
scales to 10PB. Note that this metric is related to, but different from, scalability, which is a
measure of the solution's performance as it expands.
For a compute-focused OpenStack cloud, latency of storage is a major consideration.

Using solid-state disks (SSDs) to minimize latency for instance storage and reduce CPU
delays caused by waiting for the storage will increase performance. Consider using RAID
controller cards in compute hosts to improve the performance of the underlying disk
subsystem.
The selection of storage architecture, and the corresponding storage hardware (if there is
the option), is determined by evaluating possible solutions against the key factors listed
above. This will determine if a scale-out solution (such as Ceph, GlusterFS, or similar) should
be used, or if a single, highly expandable and scalable centralized storage array would be a
better choice. If a centralized storage array is the right fit for the requirements, the hardware
will be determined by the array vendor. It is also possible to build a storage array using
commodity hardware with Open Source software, but there needs to be access to people
with expertise to build such a system. Conversely, a scale-out storage solution that uses
direct-attached storage (DAS) in the servers may be an appropriate choice. If so, then the
server hardware needs to be configured to support the storage solution.
The following lists some of the potential impacts that may affect a particular storage
architecture, and the corresponding storage hardware, of a compute-focused OpenStack
cloud:
Connectivity: Based on the storage solution selected, ensure the connectivity matches
the storage solution requirements. If a centralized storage array is selected, it is important to
determine how the hypervisors will connect to the storage array. Connectivity could affect
latency and thus performance, so check that the network characteristics will minimize
latency to boost the overall performance of the design.
Latency: Determine if the use case will have consistent or highly variable latency.
Throughput: To improve overall performance, make sure that the storage
solution throughout is optimized. While it is not likely that a compute-focused cloud will
have major data I-O to and from storage, this is an important factor to consider.
Server Hardware: If the solution uses DAS, this impacts, and is not limited to, the server
hardware choice that will ripple into host density, instance density, power density, OShypervisor, and management tools.
Where instances need to be made highly available, or they need to be capable of migration
between hosts, use of a shared storage file-system to store instance ephemeral data should
be employed to ensure that compute services can run uninterrupted in the event of a node
failure.

Selecting Networking Hardware

Some of the key considerations that should be included in the selection of networking
hardware include:
Port count: The design will require networking hardware that has the requisite port count.
Port density: The network design will be affected by the physical space that is required to
provide the requisite port count. A switch that can provide 48 10 GbE ports in 1U has a much
higher port density than a switch that provides 24 10 GbE ports in 2U. A higher port density
is preferred, as it leaves more rack space for compute or storage components that might be
required by the design. This also leads into concerns about fault domains and power density
that must also be considered. Higher density switches are more expensive and should also
be considered, as it is important not to over design the network if it is not required.
Port speed: The networking hardware must support the proposed network speed, for
example: 1 GbE, 10 GbE, or 40 GbE (or even 100 GbE).
Redundancy: The level of network hardware redundancy required is influenced by the
user requirements for high availability and cost considerations. Network redundancy can be
achieved by adding redundant power supplies or paired switches. If this is a requirement, the
hardware will need to support this configuration. User requirements will determine if a
completely redundant network infrastructure is required.
Power requirements: Ensure that the physical data center provides the necessary power
for the selected network hardware. This is not an issue for top of rack (ToR) switches, but
may be an issue for spine switches in a leaf and spine fabric, or end of row (EoR) switches.
It is important to first understand additional factors as well as the use case because these
additional factors heavily influence the cloud network architecture. Once these key
considerations have been decided, the proper network can be designed to best serve the
workloads being placed in the cloud.
It is recommended that the network architecture is designed using a scalable network model
that makes it easy to add capacity and bandwidth. A good example of such a model is the
leaf-spline model. In this type of network design, it is possible to easily add additional
bandwidth as well as scale out to additional racks of gear. It is important to select network
hardware that will support the required port count, port speed and port density while also
allowing for future growth as workload demands increase. It is also important to evaluate
where in the network architecture it is valuable to provide redundancy. Increased network
availability and redundancy comes at a cost, therefore it is recommended to weigh the cost
versus the benefit gained from utilizing and deploying redundant network switches and
using bonded interfaces at the host level.

Software Selection

Selecting software to be included in a compute-focused OpenStack architecture
design must include three main areas:
Operating system (OS) and hypervisor
OpenStack components
Supplemental software
Design decisions made in each of these areas impact the rest of the OpenStack architecture
design.

OS and Hypervisor

The selection of OS and hypervisor havs a significant impact on the end point design.
Selecting a particular operating system and hypervisor could affect server hardware
selection. For example, a selected combination needs to be supported on the
selected hardware. Ensuring the storage hardware selection and topology supports
the selected operating system and hypervisor combination should also be considered.
Additionally, make sure that the networking hardware selection and topology will work with
the chosen operating system and hypervisor combination. For example, if the design
uses Link Aggregation Control Protocol (LACP), the hypervisor needs to support it.
Some areas that could be impacted by the selection of OS and hypervisor include:
Cost: Selecting a commercially supported hypervisor such as Microsoft Hyper-V will
result in a different cost model rather than chosen a community-supported open source
hypervisor like Kinstance or Xen. Even within the ranks of open source solutions, choosing
Ubuntu over Red Hat (or vice versa) will have an impact on cost due to support contracts. On
the other hand, business or application requirements might dictate a specific
or commercially supported hypervisor.
Supportability: Depending on the selected hypervisor, the staff should have the
appropriate training and knowledge to support the selected OS and hypervisor combination.
If they do not, training will need to be provided which could have a cost impact on the
design.
Management tools: The management tools used for Ubuntu and Kinstance differ from
the management tools for VMware vSphere. Although both OS and hypervisor
combinations are supported by OpenStack, there will be very different impacts to the rest of
the design as a result of the selection of one combination versus the other.
Scale and performance: Ensure that selected OS and hypervisor combinations meet the
appropriate scale and performance requirements. The chosen architecture will need to meet

the targeted instance-host ratios with the selected OS-hypervisor combination.
Security: Ensure that the design can accommodate the regular periodic installation of
application security patches while maintaining the required workloads. The frequency of
security patches for the proposed OS - hypervisor combination will have an impact on
performance and the patch installation process could affect maintenance windows.
Supported features: Determine what features of OpenStack are required. This will often
determine the selection of the OS-hypervisor combination. Certain features are only
available with specific OSs or hypervisors. For example, if certain features are not available,
the design might need to be modified to meet the user requirements.
Interoperability: Consideration should be given to the ability of the selected OShypervisor combination to interoperate or co-exist with other OS-hypervisors, or other
software solutions in the overall design (if required). Operational and troubleshooting tools
for one OS-hypervisor combination may differ from the tools used for another OShypervisor combination and, as a result, the design will need to address if the two sets of
tools need to interoperate.

OpenStack Components

The selection of which OpenStack components will actually be included in the design and
deployed has significant impact. There are certain components that will always be present,
(Nova and Glance, for example) yet there are other services that might not need to be
present. For example, a certain design may not require OpenStack Heat. Omitting Heat
would not typically have a significant impact on the overall design. However, if the
architecture uses a replacement for OpenStack Swift for its storage component, this could
potentially have significant impacts on the rest of the design.
For a compute-focused OpenStack design architecture, the following components would be
used:
Identity (Keystone)
Dashboard (Horizon)
Compute (Nova)
Object Storage (Swift, Ceph or a commercial solution)
Image (Glance)
Networking (Neutron)
Orchestration (Heat)
OpenStack Block Storage would potentially not be incorporated into a computefocused design due to persistent block storage not being a significant requirement for the

types of workloads that would be deployed onto instances running in a compute-focused
cloud. However, there may be some situations where the need for performance dictates
that a block storage component be used to improve data I-O.
The exclusion of certain OpenStack components might also limit or constrain the
functionality of other components. If a design opts to include Heat but exclude Ceilometer,
then the design will not be able to take advantage of Heat's autoscaling functionality (which
relies on information from Ceilometer). Due to the fact that you can use Heat to spin up a
large number of instances to perform the compute-intensive processing, including Heat in a
compute-focused architecture design is strongly recommended.

Supplemental Software

While OpenStack is a fairly complete collection of software projects for building a platform
for cloud services, there are invariably additional pieces of software that might need to be
added to any given OpenStack design.

Networking Software

OpenStack Neutron provides a wide variety of networking services for instances. There are
many additional networking software packages that might be useful to manage the
OpenStack components themselves. Some examples include software to provide load
balancing, network redundancy protocols, and routing daemons. Some of these software
packages are described in more detail in the OpenStack HA Guide (refer to Chapter 8 of the
OpenStack High Availability Guide).
For a compute-focused OpenStack cloud, the OpenStack infrastructure components will
need to be highly available. If the design does not include hardware load
balancing, networking software packages like HAProxy will need to be included.

Management Software

The selected supplemental software solution impacts and affects the overall OpenStack
cloud design. This includes software for providing clustering, logging, monitoring and
alerting.
Inclusion of clustering Software, such as Corosync or Pacemaker, is determined primarily by
the availability design requirements. Therefore, the impact of including (or not including)
these software packages is primarily determined by the availability of the cloud
infrastructure and the complexity of supporting the configuration after it is deployed. The
OpenStack High Availability Guide provides more details on the installation and
configuration of Corosync and Pacemaker, should these packages need to be included in the

design.
Requirements for logging, monitoring, and alerting are determined by operational
considerations. Each of these sub-categories includes a number of various options. For
example, in the logging sub-category one might consider Logstash, Splunk, instanceware
Log Insight, or some other log aggregation-consolidation tool. Logs should be stored in a
centralized location to make it easier to perform analytics against the data. Log data
analytics engines can also provide automation and issue notification by providing a
mechanism to both alert and automatically attempt to remediate some of the more
commonly known issues.
If any of these software packages are needed, then the design must account for the
additional resource consumption (CPU, RAM, storage, and network bandwidth for a log
aggregation solution, for example). Some other potential design impacts include:
OS - Hypervisor combination: Ensure that the selected logging, monitoring, or alerting
tools support the proposed OS-hypervisor combination.
Network hardware: The network hardware selection needs to be supported by
the logging, monitoring, and alerting software.

Database Software

A large majority of the OpenStack components require access to back-end database
services to store state and configuration information. Selection of an appropriate back-end
database that will satisfy the availability and fault tolerance requirements of the OpenStack
services is required. OpenStack services support connecting to any database that is
supported by the sqlalchemy python drivers, however most common database
deployments make use of mySQL or some variation of it. It is recommended that the
database which provides back-end service within a general purpose cloud be made highly
available using an available technology which can accomplish that goal. Some of the more
common software solutions used include Galera, MariaDB and mySQL with multi-master
replication.

Prescriptive Examples

The Conseil Européen pour la Recherche Nucléaire (CERN), also known as the European
Organization for, Nuclear Research provides particle accelerators and other infrastructure
for high-energy physics research.
As of 2011 CERN operated these two compute centers in Europe with plans to add a third:
Data Center

Capacity (Approximate)

Geneva, Switzerland

3.5 Mega Watts
91000 cores
120 PB HDD
100 PB Tape
310 TB Memory

Budapest, Hungary

2.5 Mega Watts
20000 cores
6 PB HDD

To support a growing number of compute heavy users of experiments related to the Large
Hadron Collider (LHC) CERN ultimately elected to deploy an OpenStack cloud using Scientific
Linux and RDO. This effort aimed to simplify the management of the center's compute
resources with a view to doubling compute capacity through the addition of an additional
data center in 2013 while maintaining the same levels of compute staff.

Compute Architecture

The CERN solution uses Cells for segregation of compute resources and to transparently
scale between different data centers. This decision meant trading off support for security
groups and live migration. In addition some details like flavors needed to be manually
replicated across cells. In spite of these drawbacks cells were determined to provide the
required scale while exposing a single public API endpoint to users.
A compute cell was created for each of the two orginal data centers and a third was created
when a new data center was added in 2013. Each cell contains three availability zones to
further segregate compute resources and at least three RabbitMQ message brokers
configured to be clustered with mirrored queues for high availability.
The API cell, which resides behind a HAProxy load balancer, is located in the data center in

Switzerland and directs API calls to compute cells using a customized variation of the cell
scheduler. The customizations allow certain workloads to be directed to a specific data
center or "all" data centers with cell selection determined by cell RAM availability in the
latter case.

There is also some customization of the filter scheduler that handles placement within the
cells:
ImagePropertiesFilter - To provide special handling depending on the guest operating
system in use (Linux-based or Windows-based).
ProjectsToAggregateFilter - To provide special handling depending on the project the
instance is associated with.
default_schedule_zones - Allows the selection of multiple default availability zones,
rather than a single default.

The MySQL database server in each cell is managed by a central database team and
configured in an active/passive configuration with a NetApp storage backend. Backups are
performed ever 6 hours.

Network Architecture

To integrate with existing CERN networking infrastructure customizations were made to
Nova Networking. This was in the form of a driver to integrate with CERN's existing database
for tracking MAC and IP address assignments.
The driver facilitates selection of a MAC address and IP for new instances based on the
compute node the scheduler places the instance on
The driver considers the compute node that the scheduler placed an instance on and then
selects a MAC address and IP from the pre-registered list associated with that node in the
database. The database is then updated to reflect the instance the addresses were assigned
to.

Storage Architecture

The OpenStack image service is deployed in the API cell and configured to expose version 1
(V1) of the API. As a result the image registry is also required. The storage backend in use is a
3 PB Ceph cluster.
A small set of "golden" Scientific Linux 5 and 6 images are maintained which applications can
in turn be placed on using orchestration tools. Puppet is used for instance configuration
management and customization but Heat deployment is expected.

Monitoring

Although direct billing is not required, OpenStack Telemetry is used to perform metering for
the purposes of adjusting project quotas. A sharded, replicated, MongoDB backend is used.
To spread API load, instances of the nova-api service were deployed within the child cells for
Telemetry to query against. This also meant that some supporting services including
keystone, glance-api and glance-registry needed to also be configured in the child cells.

Additional monitoring tools in use include Flume (http://flume.apache.org/), Elastic Search,
Kibana (http://www.elasticsearch.org/overview/kibana/), and the CERN developed Lemon
(http://lemon.web.cern.ch/lemon/index.shtml) project.

References

The authors of the Architecture Design Guide would like to thank CERN for publicly
documenting their OpenStack deployment in these resources, which formed the basis for
this chapter:
http://openstack-in-production.blogspot.fr/
http://www.openstack.org/assets/presentation-media/Deep-Dive-into-the-CERNCloud-Infrastructure.pdf

STORAGE FOCUSED

Introduction

Cloud storage is a model of data storage where digital data is stored in logical pools
and physical storage that spans across multiple servers and locations. Cloud storage
commonly refers to a hosted object storage service, however the term has extended to
include other types of data storage that are available as a service, for example block storage.
Cloud storage is based on virtualized infrastructure and resembles broader cloud
computing in terms of accessible interfaces, elasticity, scalability, multi-tenancy,
and metered resources. Cloud storage services can be utilized from an off-premises service
or deployed on-premises.
Cloud storage is made up of many distributed, yet still synonymous resources, and is often
referred to as integrated storage clouds. Cloud storage is highly fault tolerant through
redundancy and the distribution of data. It is highly durable through the creation of
versioned copies, and can be consistent with regard to data replicas.
At a certain scale, management of data operations can become a resource intensive process
to a organization. Hierarchical storage management (HSM) systems and data grids can help
annotate and report a baseline data valuation to make intelligent decisions and
automate data decisions. HSM allows for automating tiering and movement, as well as
orchestration of data operations. A data grid is an architecture, or set of services evolving
technology, that brings together sets of services allowing users to manage large data sets.
Examples of applications that can be deployed with cloud storage characteristics are:
Active archive, backups and hierarchical storage management.
General content storage and synchronization. An example of this is private dropbox.
Data analytics with parallel file systems.
Unstructured data store for services. For example, social media backends.
Persistent block storage.
Operating system and application image store.
Media streaming.
Databases.
Content distribution.
Cloud storage peering.

User Requirements

Storage-focused clouds are defined by their requirements for data including, but not limited
to, performance, access patterns, and data structures. A balance between cost and user
requirements dictate what methods and technologies will be used in a cloud architecture.

Business Considerations

Cost: The user pays only for the storage they actually use. This limit typically reflects
average user consumption during a month. This does not mean that cloud storage is less
expensive, only that it incurs operating expenses rather the capital expenses. From a
business perspective, it should be beneficial for the solution to scale properly to prevent the
up-front purchase of a large amount of storage that goes underutilized.
Legal requirements: Multiple jurisdictions have legislative and regulatory requirements
governing the storage and management of data in cloud environments. Common areas of
regulation include data retention policies and data ownership policies.

Legal Requirements

Many jurisdictions have legislative and regulatory requirements governing the storage and
management of data in cloud environments. Common areas of regulation include:
Data retention: Policies ensuring storage of persistent data and records management to
meet data archival requirements.
Data ownership: Policies governing the possession and responsibility for data.
Data sovereignty: Policies governing the storage of data in foreign countries or otherwise
separate jurisdictions.
Data compliance: Policies governing types of information that are required to reside in
certain locations due to regular issues and cannot reside in other locations for the same
reason.
Examples of such legal frameworks include the data protection framework of the European
Union (http://ec.europa.eu/justice/data-protection/ ) and the requirements of the
Financial Industry Regulatory Authority
(http://www.finra.org/Industry/Regulation/FINRARules/ ) in the United States. Consult a
local regulatory body for more information.

Technical Requirements

The following are some technical requirements that could be incorporated into the
architecture design.

Storage Proximity: In order to provide high performance or large amounts of storage
space the design may have to accommodate storage that is each of the
hypervisors or served from a central storage device.
Performance: To boost performance the organization may want to make use of different
technologies to cache disk activity.
Availability: Specific requirements regarding availability will influence the technology
used to store and protect data. These requirements will be influence the cost and solution
that will be implemented.
Security: Data will need to be protected both in transit and at rest.

Technical Considerations

Some of the key technical considerations that are critical to a storage focused OpenStack
design architecture include:
I-O requirements: I-O performance requirements need to be researched and modeled
before deciding on a final storage framework. Running benchmarks for I-O performance will
help provide a baseline for expected performance levels. If these tests include details, for
example, object size in an object storage system or varying capacity levels for both object
storage and block storage, then the resulting data can help model behavior and results
during different workloads. Running scripted smaller benchmarks during the lifecycle of the
architecture helps record the system health at different points in time. The data from these
scripted benchmarks will assist in future scoping and gaining a deeper understanding of an
organization's needs.
Scale: The scale of the storage solution in a storage focused OpenStack architecture
design is driven both by initial requirements, including IOPS, capacity, and bandwidth, as well
as future needs. Planning capacity based on projected needs over the course of a budget
cycle is important for a design. Ideally, the chosen architecture should balance cost and
capacity while also allowing enough flexibility for new technologies and methods to be
implemented as they become available.
Security: Security design around data has multiple points of focus that vary depending on
SLAs, legal requirements, industry regulations, and certifications needed for systems or
people. HIPPA, ISO9000, and SOX compliance should be considered based on the type of
data. Levels of access control can be important for certain organizations.
OpenStack compatibility: Interoperability and integration with OpenStack can be
paramount in deciding on a storage hardware and storage management platform.
Interoperability and integration includes factors such as OpenStack Cinder interoperability,
OpenStack Swift compatibility, and hypervisor compatibility (which affects the ability to use
storage for ephemeral instance storage).
Storage management: A range of storage management-related considerations must be
addressed in the design of a storage focused OpenStack cloud. These considerations
include, but are not limited to, backup strategy (and restore strategy, since a backup that can
not be restored is useless), data valuation-hierarchical storage management, retention
strategy, data placement, and workflow automation.
Data Grids: Data grids can be helpful in deterministically answering questions around data
valuation. A fundamental challenge of today’s information sciences is determining which
data is worth keeping, on what tier of access and performance should it reside, and how long
should it remain in a storage system. Data grids improve decision making through

correlation of access patterns, ownership, and business-unit revenue with other metadata
values to deliver actionable information about data.
Strive to build a flexible design that is based on a industry standard core. One way of
accomplishing this might be through the use of different backends serving different use
cases.

Operational Considerations

Maintenance tasks: The storage solution should take into account storage maintanence
and the impact on underlying workloads.
Reliability and availability: Reliability and availablity depends on wide area network
availability and on the level of precautions taken by the service provider.
Flexibility: Organizations need to have the flexibility to choose between off-premise and
on-premise cloud storage options. This concept relies on relevant decision criteria that is
complementary to initial direct cost savings potential. For example, continuity of operations,
disaster recovery, security, and records retention laws, regulations, and policies.

Monitoring and Alerting

In a cloud environment with very high demands on storage resources, it becomes vitally
important to ensure that monitoring and alerting services have been installed and
configured to provide a real-time view into the health and performance of the storage
systems. Using an integrated management console, or other dashboards capable of
visualizing SNMP data, will be helpful in discovering and resolving issues that might arise
within the storage cluster. An example of this is Ceph’s Calamari.
A storage-focused cloud design should include:
Monitoring of physical hardware resources.
Monitoring of environmental resources such as temperature and humidity.
Monitoring of storage resources such as available storage, memory and CPU.
Monitoring of advanced storage performance data to ensure that storage systems are
performing as expected.
Monitoring of network resources for service disruptions which would affect access to
storage.
Centralized log collection.
Log analytics capabilities.
Ticketing system (or integration with a ticketing system) to track issues.
Alerting and notification of responsible teams or automated systems which will
remediate problems with storage as they arise.
Network Operations Center (NOC) staffed and always available to resolve issues.

Management Efficiency

When designing a storage solution at scale, ease of management of the storage devices is an
important consideration. Within a storage cluster, operations personnel will often be

required to replace failed drives or nodes and provide ongoing maintenance of the storage
hardware. When the cloud is designed and planned properly, the process can be performed
in a seamless and organized manner that does not have an impact on operational efficiency.
Provisioning and configuration of new or upgraded storage is another important
consideration when it comes to management of resources. The ability to easily deploy,
configure, and manage storage hardware will result in a solution which is easy to manage.
This also makes use of management systems that can automate other pieces of the overall
solution. For example, replication, retention, data backup and recovery.

Application Awareness

When designing applications that will leverage storage solutions in the cloud, design the
application to be aware of the underlying storage subsystem and the features available. As
an example, when creating an application that requires replication, it is recommended
that the application can detect whether replication is a feature natively available in the
storage subsystem. In the event that replication is not available, the application can be
designed to react in a manner such that it will be able to provide its own replication services.
An application that is designed to be aware of the underlying storage systems can be
deployed in a wide variety of infrastructures and still have the same basic
behavior regardless of the differences in the underlying infrastructure.

Fault Tolerance and Availability

Designing for fault tolerance and availability of storage systems in an OpenStack cloud is
vastly different when comparing the block storage and object storage services. The object
storage service is designed to have consistency and partition tolerance as a function of the
application. Therefore, it does not have any reliance on hardware RAID controllers to provide
redundancy for physical disks. In contrast, block storage resource nodes are commonly
configured with advanced RAID controllers and high performance disks that are designed to
provide fault tolerance at the hardware level.

Block Storage Fault Tolerance and Availability

In cases where applications require extreme performance out of block storage devices, it is
recommended to deploy high performing storage solutions such as SSD disk drives or flash
storage systems. When considering these solutions, it is important to consider the
availability of software and support to ease with the hardware and software integration
process.

In environments that place extreme demands on block storage, it is advisable to take
advantage of multiple storage pools. In this case, each pool of devices should have a similar
hardware design and disk configuration across all hardware nodes in that pool. This allows
for a design that provides applications with access to a wide variety of block storage pools,
each with their own redundancy, availability, and performance characteristics. When
deploying multiple pools of storage it is also important to consider the impact on the block
storage scheduler which is responsible for provisioning storage across resource nodes.
Ensuring that applications can schedule volumes in multiple regions, each with their own
network, power, and cooling infrastructure, can give tenants the ability to build fault tolerant
applications that will be distributed across multiple availability zones.
In addition to the block storage resource nodes, it is important to design for high availability
and redundancy of the APIs and related services that are responsible for provisioning and
providing access to storage. It is recommended to design a layer of hardware or software
load balancers in order to achieve high availability of the appropriate REST API services to
provide uninterrupted service. In some cases, it may also be necessary to deploy an
additional layer of load balancing to provide access to back-end database services
responsible for servicing and storing the state of block storage volumes. Designing a highly
available database solution to store the Cinder databases is also recommended. A number
of highly available database solutions such as Galera and MariaDB can be leveraged to help
keep database services online to provide uninterrupted access so that tenants can manage
block storage volumes.
In a cloud with extreme demands on block storage the network architecture should
take into account the amount of East-West bandwidth that will be required for instances to
make use of the available storage resources. Network devices selected should support
jumbo frames for transfering large blocks of data. In some cases, it may be necessary to
create an additional back-end storage network dedicated to providing connectivity between
instances and block storage resources so that there is no contention of network resources.

Object Storage Fault Tolerance and Availability

While consistency and partition tolerance are both inherent features of the object storage
service, it is important to design the overall storage architecture to ensure that those goals
can be met by the system being implemented. The OpenStack object storage service places
a specific number of data replicas as objects on resource nodes. These replicas are
distributed throughout the cluster based on a consistent hash ring which exists on all nodes
in the cluster.

The object storage system should be designed with sufficient number of zones to provide
quorum for the number of replicas defined. As an example, with three replicas configured in
the Swift cluster, the recommended number of zones to configure within the object storage
cluster in order to achieve quorum is 5. While it is possible to deploy a solution with fewer
zones, the implied risk of doing so is that some data may not be available and API requests
to certain objects stored in the cluster might fail. For this reason, ensure the number of
zones in the object storage cluster is properly accounted for.
Each object storage zone should be self-contained within its own availability zone. Each
availability zone should have independent access to network, power and cooling
infrastructure to ensure uninterrupted access to data. In addition, each availability zone
should be serviced by a pool of object storage proxy servers which will provide access to
data stored on the object nodes. Object proxies in each region should leverage local read and
write affinity so that access to objects is facilitated by local storage resources wherever
possible. It is recommended that upstream load balancing be deployed to ensure that proxy
services can be distributed across the multiple zones and, in some cases, it may be
necessary to make use of third party solutions to aid with geographical distribution of
services.
A zone within an object storage cluster is a logical division. A zone can be represented as a
disk within a single node, the node itself, a group of nodes in a rack, an entire rack, a data
center or even a geographic region. Deciding on the proper zone design is crucial for allowing
the object storage cluster to scale while providing an available and redundant storage
system. Furthermore, it may be necessary to configure storage policies that have different
requirements with regards to replicas, retention and other factors that could heavily affect
the design of storage in a specific zone.

Scaling Storage Services

Adding storage capacity and bandwidth is a very different process when comparing the
block and object storage services. While adding block storage capacity is a relatively simple
process, adding capacity and bandwidth to the object storage systems is a complex task that
requires careful planning and consideration during the design phase.

Scaling Block Storage

Block storage pools can be upgraded to add storage capacity rather easily without
interruption to the overall block storage service. Nodes can be added to the pool by simply
installing and configuring the appropriate hardware and software and then allowing that
node to report in to the proper storage pool via the message bus. This is because block

storage nodes report into the scheduler service advertising their availability. Once the node
is online and available tenants can make use of those storage resources instantly.
In some cases, the demand on block storage from instances may exhaust the available
network bandwidth. As a result, design network infrastructure that services block storage
resources in such a way that capacity and bandwidth can be added relatively easily. This
often involves the use of dynamic routing protocols or advanced networking solutions to
allow capacity to be added to downstream devices easily. Both the front-end and back-end
storage network designs should encompass the ability to quickly and easily add capacity
and bandwidth.

Scaling Object Storage

Adding back-end storage capacity to an object storage cluster requires careful planning and
consideration. In the design phase it is important to determine the maximum partition
power required by the object storage service, which determines the maximum number of
partitions which can exist. Object storage distributes data among all available storage, but a
partition cannot span more than one disk, so the maximum number of partitions can only be
as high as the number of disks.
For example, a system that starts with a single disk and a partion power of 3 can have 8
(2^3) partitions. Adding a second disk means that each will (usually) have 4 partitions. The
one-disk-per-partition limit means that this system can never have more than 8 disks,
limiting its scalability. However, a system that starts with a single disk and a partition power
of 10 can have up to 1024 (2^10) disks.
As back-end storage capacity is added to the system, the partition maps cause data to be
redistributed amongst storage nodes. In some cases, this replication can consist of
extremely large data sets. In these cases, it is recommended to make use of back-end
replication links which will not contend with tenants’ access to data.
As more tenants begin to access data within the cluster and their data sets grow it will
become necessary to add front-end bandwidth to service data access requests. Adding
front-end bandwidth to an object storage cluster requires careful planning and design of the
object storage proxies that will be used by tenants to gain access to the data, along with the
high availability solutions that enable easy scaling of the proxy layer. It is recommended
to design a front-end load balancing layer that tenants and consumers use to gain access to
data stored within the cluster. This load balancing layer may be distributed across zones,
regions or even across geographic boundaries, which may also require that the design

encompass geolocation solutions.
In some cases, adding bandwidth and capacity to the network resources servicing requests
between proxy servers and storage nodes will be required. For this reason, the network
architecture used for access to storage nodes and proxy servers should make use of a
design which is scalable.

Architecture

Hardware Selection

Storage hardware selection options include three areas:
Cost
Performance
Reliability
The selection of hardware for a storage-focused OpenStack cloud must reflect the fact that
the workloads are storage intensive. These workloads are not compute intensive, nor are
they consistently network intensive. The network may be heavily utilized to transfer
storage, but they are not otherwise network intensive. The hardware selection for a storagefocused OpenStack architecture design must reflect this preference for storage-intensive
workloads.

Storage Hardware Selection

For a storage-focused OpenStack design architecture, the selection of storage hardware will
determine the overall performance and scalability of the design architecture. A number of
different factors must be considered in the design process:
Cost: The overall cost of the solution will play a major role in what storage architecture
and the resulting storage hardware that is selected.
Performance: The performance of the solution, measured by observing the latency of
storage I-O requests, also plays a major role. In a compute-focused OpenStack
cloud storage latency could potentially be a major consideration, in some computeintensive workloads, minimizing the delays that the CPU experiences while fetching data
from the storage can have a significant impact on the overall performance of the
application.
Scalability: "Scalability" refers to how well the storage solution performs as it is expanded
up to its maximum size. A storage solution that performs well in small configurations but
has degrading performance as it expands would be considered not scalable. Conversely, a
solution that continues to perform well at maximum expansion would be considered
scalable. The ability of the storage solution to continue to perform well as it expands is
important.
Expandability: Here we are referring to the overall ability of the solution to grow. A
storage solution that expands to 50 PB is considered more expandable than a solution that
only scales to 10 PB. Note that this metric is related to but different from scalability which is
a measure of the solution's performance as it expands.

Latency is one of the key considerations in a storage-focused OpenStack cloud . Using solidstate disks (SSDs) to minimize latency for instance storage and reduce CPU delays caused by
waiting for the storage will have a result of increased performance. It is also recommended
to evaluate the gains from using RAID controller cards in compute hosts to improve the
performance of the underlying disk subsystem.
The selection of storage architecture (and the corresponding storage hardware, if there is an
option) is determined by evaluating possible solutions against the key factors above. This
will determine if a scale-out solution (such as Ceph, GlusterFS, or similar) should be used or
if a single, highly expandable and scalable centralized storage array would be a better choice.
If a centralized storage array is the right fit for the requirements then the hardware will be
determined by the array vendor. It is possible to build a storage array using commodity
hardware with Open Source software, but there needs to be access to people with expertise
to build such a system. On the other hand, a scale-out storage solution that uses directattached storage (DAS) in the servers may be an appropriate choice. If this is true, then the
server hardware needs to be configured to support the storage solution.
Some potential impacts that might affect a particular storage architecture (and
corresponding storage hardware) of a Storage-focused OpenStack cloud:
Connectivity: Based on the storage solution selected, ensure the connectivity matches
the storage solution requirements. If a centralized storage array is selected it is important to
determine how the hypervisors will connect to the storage array. Connectivity can affect
latency and thus performance. It is recommended to check that the network characteristics
will minimize latency to boost the overall performance of the design.
Latency: Determine if the use case will have consistent or highly variable latency.
Throughput: Ensure that the storage solution throughput is optimized based on
application requirements.
Server Hardware: Use of DAS impacts the server hardware choice and affects host
density, instance density, power density, OS-hypervisor, and management tools, to name a
few.

Compute (Server) Hardware Selection

Compute (server) hardware must be evaluated against four opposing dimensions:
Server density: A measure of how many servers can fit into a given measure of physical
space, such as a rack unit [U].
Resource capacity: The number of CPU cores, how much RAM, or how much storage a
given server will deliver.

Expandability: The number of additional resources that can be added to a server before it
has reached its limit.
Cost: The relative of the hardware weighted against the level of design effort needed to
build the system.
The dimensions need to be weighed against each other to determine the best design for the
desired purpose. For example, increasing server density can mean sacrificing resource
capacity or expandability. Increasing resource capacity and expandability can increase cost
but decrease server density. Decreasing cost often means decreasing supportability, server
density, resource capacity, and expandability.
For a storage-focused OpenStack architecture design, a secondary design consideration for
selecting server hardware will be the compute capacity (CPU cores and RAM capacity). As a
result, the required server hardware must supply adequate CPU sockets, additional CPU
cores, and more RAM; network connectivity and storage capacity are not as critical. The
hardware will need to provide enough network connectivity and storage capacity to meet
the user requirements, however they are not the primary consideration.
Since there is only a need for adequate CPU and RAM capacity, some server hardware form
factors will be better suited to this storage-focused design than others:
Most blade servers typically support dual-socket multi-core CPUs; to avoid the limit will
mean choosing "full width" or "full height" blades, which means losing server density. The
high density blade servers (for example, both HP BladeSystem and Dell PowerEdge
M1000e), which support up to 16 servers in only 10 rack units using "half height" or "half
width" blades, suddenly decrease the density by 50% (only 8 servers in 10 U) if a "full width"
or "full height" option is used.
1U rack-mounted servers (servers that occupy only a single rack unit) might be able to
offer greater server density than a blade server solution (40 servers in a rack, providing
space for the top of rack (ToR) switches, versus 32 "full width" or "full height" blade servers
in a rack), but often are limited to dual-socket, multi-core CPU configurations. Note that as
of the Icehouse release, neither HP, IBM, nor Dell offered 1U rack servers with more than 2
CPU sockets. To obtain greater than dual-socket support in a 1U rack-mount form factor,
customers need to buy their systems from Original Design Manufacturers (ODMs) or
second-tier manufacturers. This may cause issues for organizations that have preferred
vendor policies or concerns with support and hardware warranties of non-tier 1 vendors.
2U rack-mounted servers provide quad-socket, multi-core CPU support but with a
corresponding decrease in server density (half the density offered by 1U rack-mounted
servers).

Larger rack-mounted servers, such as 4U servers, often provide even greater CPU
capacity. Commonly supporting four or even eight CPU sockets. These servers have greater
expandability capacity but such servers have much lower server density and usually greater
hardware cost.
The so-called "sled servers" (rack-mounted servers that support multiple independent
servers in a single 2U or 3U enclosure) deliver increased density as compared to a typical
1U-2U rack-mounted servers. For example, many sled servers offer four independent dualsocket nodes in 2U for a total of 8 CPU sockets in 2U. However, the dual-socket limitation
on individual nodes may not be sufficient to offset their additional cost and configuration
complexity.
Other factors that will strongly influence server hardware selection for a storage-focused
OpenStack design architecture:
Instance density: In this architecture, instance density and CPU-RAM oversubscription
are lower. More hosts will be required to support the anticipated scale, especially if the
design uses dual-socket hardware designs.
Host density: Another option to address the higher host count is to use a quad socket
platform. Taking this approach will decrease host density which also increases rack
count. This configuration affects the number of power connections and
also impacts network and cooling requirements.
Power and cooling density: The power and cooling density requirements might be lower
than with blade, sled, or 1U server designs due to lower host density (by using 2U, 3U or
even 4U server designs). For data centers with older infrastructure, this might be a desirable
feature.
Storage-focused OpenStack design architecture server hardware selection should focus
on a "scale up" versus "scale out" solution. The determination of which will be the best
solution, smaller number of larger hosts or a larger number of smaller hosts, will depend of
a combination of factors including cost, power, cooling, physical rack and floor space,
support-warranty, and manageability.

Networking Hardware Selection

Some of the key considerations that should be included in the selection of networking
hardware include:
Port count: The user will require networking hardware that has the requisite port count.
Port density: The network design will be affected by the physical space that is required to
provide the requisite port count. A switch that can provide 48 10 GbE ports in 1U has a much
higher port density than a switch that provides 24 10 GbE ports in 2U. On a general scale,

a higher port density leaves more rack space for compute or storage components which is
preferred. It is also important to consider fault domains and power density. Finally, higher
density switches are more expensive, therefore it is important not to over design the
network.
Port speed: The networking hardware must support the proposed network speed, for
example: 1 GbE, 10 GbE, or 40 GbE (or even 100 GbE).
Redundancy: The level of network hardware redundancy required is influenced by the
user requirements for high availability and cost considerations. Network redundancy can be
achieved by adding redundant power supplies or paired switches. If this is a requirement the
hardware will need to support this configuration. User requirements will determine if a
completely redundant network infrastructure is required.
Power requirements: Make sure that the physical data center provides the necessary
power for the selected network hardware. This is not typically an issue for top of rack (ToR)
switches, but may be an issue for spine switches in a leaf and spine fabric, or end of row
(EoR) switches.
Protocol support: It is possible to gain even more performance out of a single storage
system by using specialized network technologies such as RDMA, SRP, iSER and SCST. The
specifics for using these technologies is beyond the scope of this book.

Software Selection

Selecting software to be included in a storage-focused OpenStack architecture design
includes three areas:
Operating system (OS) and hypervisor
OpenStack components
Supplemental software
Design decisions made in each of these areas impacts the rest of the OpenStack architecture
design.

Operating System and Hypervisor

The selection of OS and hypervisor has a significant impact on the overall design and also
affects server hardware selection. Ensure that the storage hardware is supported by
the selected operating system and hypervisor combination and that the networking
hardware selection and topology will work with the chosen operating system
and hypervisor combination. For example, if the design uses Link Aggregation Control
Protocol (LACP), the OS and hypervisor are both required to support it.
Some areas that could be impacted by the selection of OS and hypervisor include:

Cost: Selection of a commercially supported hypervisor, such as Microsoft Hyper-V, will
result in a different cost model rather than selected a community-supported open source
hypervisor like Kinstance or Xen. Similarly, choosing Ubuntu over Red Hat (or vice versa) will
have an impact on cost due to support contracts. Conversely, business or application
requirements might dictate a specific or commercially supported hypervisor.
Supportability: Whichever hypervisor is chosen, the staff needs to have appropriate
training and knowledge to support the selected OS and hypervisor combination. If they do
not training will need to be provided, which could have a cost impact on the design.
Another aspect to consider would be the support for the OS-hypervisor. The support of a
commercial product such as Redhat, Suse, or Windows, is the responsibility of the OS
vendor. If an Open Source platform is chosen, the support comes from in-house resources.
Either decision has a cost that will have an impact on the design.
Management tools: The management tools used for Ubuntu and Kinstance differ
from the management tools for VMware vSphere. Although both OS and hypervisor
combinations are supported by OpenStack, there will naturally be very different impacts to
the rest of the design as a result of the selection of one combination versus the other.
Scale and performance: Make sure that selected OS and hypervisor combination meet
the appropriate scale and performance requirements needed for this general purpose
OpenStack cloud. The chosen architecture will need to meet the targeted instance-host
ratios with the selected OS-hypervisor combination.
Security: Make sure that the design can accommodate the regular periodic installation of
application security patches while maintaining the required workloads. The frequency of
security patches for the proposed OS-hypervisor combination will have an impact on
performance and the patch installation process could affect maintenance windows.
Supported features: Determine what features of OpenStack are required. This will often
determine the selection of the OS-hypervisor combination. Certain features are only
available with specific OSs or hypervisors. For example, if certain features are not available,
the design might need to be modified to meet the user requirements.
Interoperability: Consideration should be given to the ability of the selected OShypervisor combination to interoperate or co-exist with other OS-hypervisors ,or other
software solutions in the overall design, if that is a requirement. Operational
and troubleshooting tools for one OS-hypervisor combination may differ from the tools
used for another OS-hypervisor combination. As a result, the design will need to address if
the two sets of tools need to interoperate.

OpenStack Components

The selection of OpenStack components has a significant direct impact on the overall

design. While there are certain components that will always be present, (Nova and Glance,
for example) there are other services that may not need to be present. As an example, a
certain design may not require OpenStack Heat. Omitting Heat would not typically have a
significant impact on the overall design however, if the architecture uses a replacement
for OpenStack Swift for its storage component, this could potentially have significant
impacts on the rest of the design.
A storage-focused design might require the ability to use Heat to launch instances with
Cinder volumes to perform storage-intensive processing.
For a storage-focused OpenStack design architecture, the following components would
typically be used:
Keystone
Horizon
Nova (including the use of multiple hypervisor drivers)
Swift (or another object storage solution)
Cinder
Glance
Neutron or nova-network
The exclusion of certain OpenStack components may limit or constrain the functionality of
other components. If a design opts to include Heat but exclude Ceilometer, then the design
will not be able to take advantage of Heat's auto scaling functionality (which relies on
information from Ceilometer). Due to the fact that you can use Heat to spin up a large
number of instances to perform the compute-intensive processing, including Heat in a
compute-focused architecture design is strongly recommended.

Supplemental Software

While OpenStack is a fairly complete collection of software projects for building a platform
for cloud services, there are additional pieces of software that might need to be added to
any given OpenStack design.

Networking Software

OpenStack Neutron provides a wide variety of networking services for instances. There are
many additional networking software packages that may be useful to manage the
OpenStack components themselves. Some examples include HAProxy, keepalived, and
various routing daemons (like Quagga). Some of these software packages, HAProxy in
particular, are described in more detail in the OpenStack HA Guide (refer to Chapter 8 of the

OpenStack High Availability Guide). For a general purpose OpenStack cloud, it is reasonably
likely that the OpenStack infrastructure components will need to be highly available, and
therefore networking software packages like HAProxy will need to be included.

Management Software

This includes software for providing clustering, logging, monitoring, and alerting. The factors
for determining which software packages in this category should be selected is outside the
scope of this design guide. This design guide focuses specifically on how the selected
supplemental software solution impacts or affects the overall OpenStack cloud design.
Clustering Software, such as Corosync or Pacemaker, is determined primarily by the
availability design requirements. Therefore, the impact of including (or not including) these
software packages is determined by the availability of the cloud infrastructure and the
complexity of supporting the configuration after it is deployed. The OpenStack High
Availability Guide provides more details on the installation and configuration of Corosync
and Pacemaker, should these packages need to be included in the design.
Requirements for logging, monitoring, and alerting are determined by operational
considerations. Each of these sub-categories includes a number of various options. For
example, in the logging sub-category one might consider Logstash, Splunk, instanceware
Log Insight, or some other log aggregation-consolidation tool. Logs should be stored in a
centralized location to make it easier to perform analytics against the data. Log data
analytics engines can also provide automation and issue notification, by providing a
mechanism to both alert and automatically attempt to remediate some of the more
commonly known issues.
If any of these software packages are needed, then the design must account for the
additional resource consumption (CPU, RAM, storage, and network bandwidth for a log
aggregation solution, for example). Some other potential design impacts include:
OS - Hypervisor combination: Ensure that the selected logging, monitoring, or alerting
tools support the proposed OS-hypervisor combination.
Network hardware: The network hardware selection needs to be supported by
the logging, monitoring, and alerting software.

Database Software

Virtually all of the OpenStack components require access to back-end database services to
store state and configuration information. Choose an appropriate back-end database which
will satisfy the availability and fault tolerance requirements of the OpenStack services.

MySQL is generally considered to be the de facto database for OpenStack, however, other
compatible databases are also known to work. Note, however, that Ceilometer uses
MongoDB.
The solution selected to provide high availability for the database will change based on the
selected database. If MySQL is selected, then a number of options are available. For activeactive clustering a replication technology such as Galera can be used. For active-passive
some form of shared storage must be used. Each of these potential solutions has an impact
on the design:
Solutions that employ Galera/MariaDB will require at least three MySQL nodes.
MongoDB will have its own design considerations, with regards to making the database
highly available.
OpenStack design, generally, does not include shared storage but for a high availability
design some components might require it depending on the specific implementation.

Prescriptive Examples

Storage-focused architectures are highly dependant on the specific use case. Three specific
example use cases are discussed in this section: an object store with a RESTful
interface, compute analytics with parallel file systems, and a high performance database.

Object Store with REST Interface

This example describes a REST interface without a high performance requirement, so the
presented REST interface does not require a high performance caching tier, and is presented
as a traditional Object store running on traditional spindles.
Swift is a highly scalable object store that is part of the OpenStack project. This is a diagram
to explain the example architecture:

This example uses the following components:
Network:
10 GbE horizontally scalable spine leaf backend storage and frontend network.

Storage hardware:
10 storage servers each with 12x4 TB disks which equals 480 TB total space with
approximately 160 Tb of usable space after replicas.
Proxy:
3x proxies
2x10 GbE bonded frontend
2x10 GbE backend bonds
Approximately 60 Gb of total bandwidth to the backend storage cluster
NOTE: For some applications, it may be necessary to implement a 3rd party caching layer to
achieve suitable performance.

Compute Analytics with Sahara

Analytics of large data sets can be highly dependant on the performance of the storage
system. Some clouds using storage systems such as HDFS have inefficiencies which can
cause performance issues. A potential solution to this is to implement a storage system
designed with performance in mind. Traditionally, parallel file systems have filled this need
in the HPC space and could be a consideration, when applicable, for large scale
performance-oriented systems.
This example discusses an OpenStack Object Store with a high performance requirement.
OpenStack has integration with Hadoop through the Sahara project, which is leveraged to
manage the Hadoop cluster within the cloud.

The actual hardware requirements and configuration are similar to those of the High
Performance Database example below. In this case, the architecture uses Ceph's Swiftcompatible REST interface, features that allow for connecting a caching pool to allow for
acceleration of the presented pool.

High Performance Database with Trove

Databases are a common workload that can greatly benefit from a high performance
storage backend. Although enterprise storage is not a requirement, many environments
have existing storage that can be used as backends for an OpenStack cloud. As shown in the
following diagram, a storage pool can be carved up to provide block devices with OpenStack
Block Storage to instances as well as an object interface. In this example the database I-O
requirements were high and demanded storage presented from a fast SSD pool.
A storage system is used to present a LUN that is backed by a set of SSDs using a traditional

storage array with OpenStack Block Storage integration or a storage platform such as Ceph
or Gluster.
This kind of system can also provide additional performance in other situations. For
example, in the database example below, a portion of the SSD pool can act as a block device
to the Database server. In the high performance analytics example, the REST interface
would be accelerated by the inline SSD cache layer.

Ceph was selected to present a Swift compatible REST interface, as well as a block level
storage from a distributed storage cluster. It is highly flexible and has features that allow to
reduce cost of operations such as self healing and auto balancing. Erasure coded pools are
used to maximize the amount of usable space. Note that there are special considerations
around erasure coded pools, for example, higher computational requirements and

limitations on the operations allowed on a object. For example, partial writes are not
supported in an erasure coded pool.
A potential architecture for Ceph, as it relates to the examples above, would entail the
following:
Network:
10 GbE horizontally scalable spine leaf backend storage and frontend network
Storage hardware:
5 storage servers for caching layer 24x1 TB SSD
10 storage servers each with 12x4 TB disks which equals 480 TB total space with about
approximately 160 Tb of usable space after 3 replicas
REST proxy:
3x proxies
2x10 GbE bonded frontend
2x10 GbE backend bonds
Approximately 60 Gb of total bandwidth to the backend storage cluster
The SSD cache layer is used to present block devices directly to Hypervisors or
instances. The SSD cache systems can also be used as an inline cache for the REST
interface.

NETWORK FOCUSED

Introduction

All OpenStack deployments are dependent, to some extent, on network communication in
order to function properly due to a service-based nature. In some cases, however, use cases
dictate that the network is elevated beyond simple infrastructure. This section is a
discussion of architectures that are more reliant or focused on network services. These
architectures are heavily dependent on the network infrastructure and need to be
architected so that the network services perform and are reliable in order to satisfy user and
application requirements.
Some possible use cases include:
Content Delivery Network: This could include streaming video, photographs or any other
cloud based repository of data that is distributed to a large number of end users. Mass
market streaming video will be very heavily affected by the network configurations that
would affect latency, bandwidth, and the distribution of instances. Not all video streaming is
consumer focused. For example, multicast videos (used for media, press conferences,
corporate presentations, web conferencing services, etc.) can also utilize a content delivery
network. Content delivery will be affected by the location of the video repository and its
relationship to end users. Performance is also affected by network throughput of the
backend systems, as well as the WAN architecture and the cache methodology.
Network Management Functions: A cloud that provides network service functions would
be built to support the delivery of backend network services such as DNS, NTP or SNMP and
would be used by a company for internal network management.
Network Service Offerings: A cloud can be used to run customer facing network tools to
support services. For example, VPNs, MPLS private networks, GRE tunnels and others.
Web portals / Web Services: Web servers are a common application for cloud
services and it is recommended to have an understanding of the network requirements. The
network will need to be able to scale out to meet user demand and deliver webpages with a
minimum of latency. Internal east-west and north-south network bandwidth must be
considered depending on the details of the portal architecture.
High Speed and High Volume Transactional Systems: These types of applications are
very sensitive to network configurations. Examples include many financial systems, credit
card transaction applications, trading and other extremely high volume systems. These
systems are sensitive to network jitter and latency. They also have a high volume of both
east-west and north-south network traffic that needs to be balanced to maximize efficiency
of the data delivery. Many of these systems have large high performance database
backends that need to be accessed.

High Availability: These types of use cases are highly dependent on the proper sizing of
the network to maintain replication of data between sites for high availability. If one site
becomes unavailable, the extra sites will be able to serve the displaced load until the original
site returns to service. It is important to size network capacity to handle the loads that are
desired.
Big Data: Clouds that will be used for the management and collection of big data (data
ingest) will have a significant demand on network resources. Big data often uses partial
replicas of the data to maintain data integrity over large distributed clouds. Other big data
applications that require a large amount of network resources are Hadoop, Cassandra,
NuoDB, RIAK and other No-SQL and distributed databases.
Virtual Desktop Infrastructure (VDI): This use case is very sensitive to network
congestion, latency, jitter and other network characteristics. Like video streaming, the user
experience is very important however, unlike video streaming, caching is not an option to
offset the network issues. VDI requires both upstream and downstream traffic and cannot
rely on caching for the delivery of the application to the end user.
Voice over IP (VoIP): This is extremely sensitive to network congestion, latency, jitter and
other network characteristics. VoIP has a symmetrical traffic pattern and it requires network
quality of service (QoS) for best performance. It may also require an active queue
management implementation to ensure delivery. Users are very sensitive to latency and
jitter fluctuations and can detect them at very low levels.
Video Conference / Web Conference: This also is extremely sensitive to network
congestion, latency, jitter and other network flaws. Video Conferencing has a symmetrical
traffic pattern, but unless the network is on an MPLS private network, it cannot use network
quality of service (QoS) to improve performance. Similar to VOIP, users will be sensitive to
network performance issues even at low levels.
High Performance Computing (HPC): This is a complex use case that requires careful
consideration of the traffic flows and usage patterns to address the needs of cloud clusters.
It has high East-West traffic patterns for distributed computing, but there can be
substantial North-South traffic depending on the specific application.

User Requirements

Network focused architectures vary from the general purpose designs. They are heavily
influenced by a specific subset of applications that interact with the network in a more
impacting way. Some of the business requirements that will influence the design include:
User experience: User experience is impacted by network latency through slow page
loads, degraded video streams, and low quality VoIP sessions. Users are often not aware of
how network design and architecture affects their experiences. Both enterprise customers
and end-users rely on the network for delivery of an application. Network performance
problems can provide a negative experience for the end-user, as well as productivity and
economic loss.
Regulatory requirements: Networks need to take into consideration any regulatory
requirements about the physical location of data as it traverses the network. For example,
Canadian medical records cannot pass outside of Canadian sovereign territory. Another
network consideration is maintaining network segregation of private data flows and
ensuring that the network between cloud locations is encrypted where required. Network
architectures are affected by regulatory requirements for encryption and protection of data
in flight as the data moves through various networks.

Legal Requirements

Many jurisdictions have legislative and regulatory requirements governing the storage and
management of data in cloud environments. Common areas of regulation include:
Data retention policies ensuring storage of persistent data and records management to
meet data archival requirements.
Data ownership policies governing the possession and responsibility for data.
Data sovereignty policies governing the storage of data in foreign countries or otherwise
separate jurisdictions.
Data compliance policies governing where information needs to reside in certain locations
due to regular issues and, more importantly, where it cannot reside in other locations for the
same reason.
Examples of such legal frameworks include the data protection framework of the European
Union (http://ec.europa.eu/justice/data-protection/ ) and the requirements of the Financial
Industry Regulatory Authority (http://www.finra.org/Industry/Regulation/FINRARules) in
the United States. Consult a local regulatory body for more information.

High Availability Issues

OpenStack installations with high demand on network resources have high
availability requirements that are determined by the application and use
case. Financial transaction systems will have a much higher requirement for high availability
than a development application. Forms of network availability, for example quality of
service (QoS), can be used to improve the network performance of sensitive applications, for
example VoIP and video streaming.
Often, high performance systems will have SLA requirements for a minimum QoS with
regard to guaranteed uptime, latency and bandwidth. The level of the SLA can have a
significant impact on the network architecture and requirements for redundancy in the
systems.

Risks

Network Misconfigurations: Configuring incorrect IP addresses, VLANs, and routes can
cause outages to areas of the network or, in the worst case scenario, the entire cloud
infrastructure. Misconfigurations can cause disruptive problems and should be automated
to minimize the opportunity for operator error.
Capacity Planning: Cloud networks need to be managed for capacity and growth over
time. There is a risk that the network will not grow to support the workload. Capacity
planning includes the purchase of network circuits and hardware that can potentially have
lead times measured in months or more.
Network Tuning: Cloud networks need to be configured to minimize link loss, packet loss,
packet storms, broadcast storms, and loops.
Single Point Of Failure (SPOF): High availability must be taken into account even at the
physical and environmental layers. If there is a single point of failure due to only one
upstream link, or only one power supply, an outage becomes unavoidable.
Complexity: An overly complex network design becomes difficult to maintain and
troubleshoot. While automated tools that handle overlay networks or device level
configuration can mitigate this, non-traditional interconnects between functions and
specialized hardware need to be well documented or avoided to prevent outages.
Non-standard features: There are additional risks that arise from configuring the cloud
network to take advantage of vendor specific features. One example is multi-link
aggregation (MLAG) that is being used to provide redundancy at the aggregator switch level
of the network. MLAG is not a standard and, as a result, each vendor has their own
proprietary implementation of the feature. MLAG architectures are not interoperable across
switch vendors, which leads to vendor lock-in, and can cause delays or inability when
upgrading components.

Security

Security is often overlooked or added after a design has been implemented. Consider
security implications and requirements before designing the physical and logical network
topologies. Some of the factors that need to be addressed include making sure the
networks are properly segregated and traffic flows are going to the correct
destinations without crossing through locations that are undesirable. Some examples of
factors that need to be taken into consideration are:
Firewalls
Overlay interconnects for joining separated tenant networks
Routing through or avoiding specific networks
Another security vulnerability that must be taken into account is how networks are attached
to hypervisors. If a network must be separated from other systems at all costs, it may be
necessary to schedule instances for that network onto dedicated compute nodes. This may
also be done to mitigate against exploiting a hypervisor breakout allowing the attacker
access to networks from a compromised instance.

Technical Considerations

Designing an OpenStack network architecture involves a combination of layer 2 and layer 3
considerations. Layer 2 decisions involve those made at the data-link layer, such as the
decision to use Ethernet versus Token Ring. Layer 3 involve those made about the protocol
layer and the point at which IP comes into the picture. As an example, a completely internal
OpenStack network can exist at layer 2 and ignore layer 3 however, in order for any traffic to
go outside of that cloud, to another network, or to the Internet, a layer 3 router or switch
must be involved.

Network Architecture: Comparing Layer 2 and Layer 3 Options

The past few years have seen two competing trends in networking. There has been a trend
towards building data center network architectures based on layer 2 networking and
simultaneously another network architecture approach is to treat the cloud environment
essentially as a miniature version of the Internet. This represents a radically different
approach to the network architecture from what is currently installed in the staging
environment because the Internet is based entirely on layer 3 routing rather than layer 2
switching.

Ethernet Frames: Layer 2 Architecture Advantages

In the data center context, there are advantages of designing the network on layer 2
protocols rather than layer 3. In spite of the difficulties of using a bridge to perform the
network role of a router, many vendors, customers, and service providers are attracted to
the idea of using Ethernet in as many parts of their networks as possible. The benefits of
selecting a layer 2 design are:
Ethernet frames contain all the essentials for networking. These include, but are not
limited to, globally unique source addresses, globally unique destination addresses, and
error control.
Ethernet frames can carry any kind of packet. Networking at layer 2 is independent of the
layer 3 protocol.
More layers added to the Ethernet frame only slow the networking process down. This is
known as 'nodal processing delay'.
Adjunct networking features, for example class of service (CoS) or multicasting, can be
added to Ethernet as readily as IP networks.
VLANs are an easy mechanism for isolating networks.
Most information starts and ends inside Ethernet frames. Today this applies to data, voice
(for example, VoIP) and video (for example, web cameras). The concept is that, if more of the

end-to-end transfer of information from a source to a destination can be done in the form
of Ethernet frames, more of the benefits of Ethernet can be realized on the network. Though
it is not a substitute for IP networking, networking at layer 2 can be a powerful adjunct to IP
networking.
The basic reasoning behind using layer 2 Ethernet over layer 3 IP networks is the speed, the
reduced overhead of the IP hierarchy, and the lack of requirement to keep track of IP
address configuration as systems are moved around. Whereas the simplicity of layer 2
protocols might work well in a data center with hundreds of physical machines, cloud data
centers have the additional burden of needing to keep track of all virtual machine addresses
and networks. In these data centers, it is not uncommon for one physical node to support
30-40 instances.
Important Note: Networking at the frame level says nothing about the presence or absence
of IP addresses at the packet level. Almost all ports, links, and devices on a network of LAN
switches still have IP addresses, as do all the source and destination hosts. There are many
reasons for the continued need for IP addressing. The largest one is the need to manage the
network. A device or link without an IP address is usually invisible to most management
applications. Utilities including remote access for diagnostics, file transfer of configurations
and software, and similar applications cannot run without IP addresses as well as MAC
addresses.

Layer 2 Architecture Limitations

Outside of the traditional data center the limitations of layer 2 network architectures
become more obvious.
Number of VLANs is limited to 4096.
The number of MACs stored in switch tables is limited.
The need to maintain a set of layer 4 devices to handle traffic control must be
accommodated.
MLAG, often used for switch redundancy, is a proprietary solution that does not scale
beyond two devices and forces vendor lock-in.
It can be difficult to troubleshoot a network without IP addresses and ICMP.
Configuring ARP is considered complicated on large layer 2 networks.
All network devices need to be aware of all MACs, even instance MACs, so there is
constant churn in MAC tables and network state changes as instances are started or
stopped.
Migrating MACs (instance migration) to different physical locations are a
potential problem if ARP table timeouts are not set properly.

It is important to know that layer 2 has a very limited set of network management tools. It is
very difficult to control traffic, as it does not have mechanisms to manage the network or
shape the traffic, and network troubleshooting is very difficult. One reason for this
difficulty is network devices have no IP addresses. As a result, there is no reasonable way to
check network delay in a layer 2 network.
On large layer 2 networks, configuring ARP learning can also be complicated. The setting for
the MAC address timer on switches is critical and, if set incorrectly, can cause significant
performance problems. As an example, the Cisco default MAC address timer is extremely
long. Migrating MACs to different physical locations to support instance migration can be a
significant problem. In this case, the network information maintained in the switches could
be out of sync with the new location of the instance.
In a layer 2 network, all devices are aware of all MACs, even those that belong to instances.
The network state information in the backbone changes whenever an instance is started or
stopped. As a result there is far too much churn in the MAC tables on the backbone
switches.

Layer 3 Architecture Advantages

In the layer 3 case, there is no churn in the routing tables due to instances starting and
stopping. The only time there would be a routing state change would be in the case of a Top
of Rack (ToR) switch failure or a link failure in the backbone itself. Other advantages of using
a layer 3 architecture include:
layer 3 networks provide the same level of resiliency and scalability as the Internet.
Controlling traffic with routing metrics is straightforward.
layer 3 can be configured to use BGP confederation for scalability so core routers have
state proportional to number of racks, not to the number of servers or instances.
Routing ensures that instance MAC and IP addresses out of the network core reducing
state churn. Routing state changes only occur in the case of a ToR switch failure or
backbone link failure.
There are a variety of well tested tools, for example ICMP, to monitor and manage traffic.
layer 3 architectures allow for the use of Quality of Service (QoS) to manage network
performance.

Layer 3 Architecture Limitations

The main limitation of layer 3 is that there is no built-in isolation mechanism comparable to
the VLANs in layer 2 networks. Furthermore, the hierarchical nature of IP addresses means
that an instance will also be on the same subnet as its physical host. This means that it

cannot be migrated outside of the subnet easily. For these reasons, network virtualization
needs to use IP encapsulation and software at the end hosts for both isolation, as well as for
separation of the addressing in the virtual layer from addressing in the physical layer. Other
potential disadvantages of layer 3 include the need to design an IP addressing scheme
rather than relying on the switches to automatically keep track of the MAC addresses and to
configure the interior gateway routing protocol in the switches.

Network Recommendations Overview

OpenStack has complex networking requirements for several reasons. Many components
interact at different levels of the system stack that adds complexity. Data flows are
complex. Data in an OpenStack cloud moves both between instances across the network
(also known as East-West), as well as in and out of the system (also known as North-South).
Physical server nodes have network requirements that are independent of those used
by instances which need to be isolated from the core network to account for scalability. It is
also recommended to functionally separate the networks for security purposes and tune
performance through traffic shaping.
A number of important general technical and business factors need to be taken into
consideration when planning and designing an OpenStack network. They include:
A requirement for vendor independence. To avoid hardware or software vendor lock-in,
the design should not rely on specific features of a vendor’s router or switch.
A requirement to massively scale the ecosystem to support millions of end users.
A requirement to support indeterminate platforms and applications.
A requirement to design for cost efficient operations to take advantage of massive scale.
A requirement to ensure that there is no single point of failure in the cloud ecosystem.
A requirement for high availability architecture to meet customer SLA requirements.
A requirement to be tolerant of rack level failure.
A requirement to maximize flexibility to architect future production environments.
Keeping all of these in mind, the following network design recommendations can be made:
Layer 3 designs are preferred over layer 2 architectures.
Design a dense multi-path network core to support multi-directional scaling and
flexibility.
Use hierarchical addressing because it is the only viable option to scale network
ecosystem.
Use virtual networking to isolate instance service network traffic from the management
and internal network traffic.

Isolate virtual networks using encapsulation technologies.
Use traffic shaping for performance tuning.
Use eBGP to connect to the Internet up-link.
Use iBGP to flatten the internal traffic on the layer 3 mesh.
Determine the most effective configuration for block storage network.

Additional Considerations

There are numerous topics to consider when designing a network-focused OpenStack
cloud.

OpenStack Networking versus Nova Network Considerations

Selecting the type of networking technology to implement depends on many factors.
OpenStack Networking (Neutron) and Nova Network both have their advantages and
disadvantages. They are both valid and supported options that fit different use cases as
described in the following table.
OpenStack Networking versus Nova Network
Nova Network
Simple, single agent
More mature, established
Flat or VLAN
No plugin support
Scales well
No multi-tier topologies

OpenStack Networking
Complex, multiple agents
Newer, maturing
Flat, VLAN, Overlays, L2-L3, SDN
Plugin support for 3rd parties
Scaling requires 3rd party plugins
Multi-tier topologies

Redundant Networking: ToR Switch High Availability Risk Analysis

A technical consideration of networking is the idea that switching gear in the data center that
should be installed with backup switches in case of hardware failure.
Research into the mean time between failures (MTBF) on switches is between 100,000 and
200,000 hours. This number is dependent on the ambient temperature of the switch in the
data center. When properly cooled and maintained, this translates to between 11 and 22
years before failure. Even in the worst case of poor ventilation and high ambient
temperatures in the data center, the MTBF is still 2-3 years. This is based on published
research found at
http://www.garrettcom.com/techsupport/papers/ethernet_switch_reliability.pdf and
http://www.n-tron.com/pdf/network_availability.pdf
In most cases, it is much more economical to only use a single switch with a small pool of

spare switches to replace failed units than it is to outfit an entire data center with redundant
switches. Applications should also be able to tolerate rack level outages without affecting
normal operations since network and compute resources are easily provisioned and
plentiful.

Preparing for the future: IPv6 Support

One of the most important networking topics today is the impending exhaustion of IPv4
addresses. In early 2014, ICANN announced that they started allocating the final IPv4
address blocks to the Regional Internet Registries
http://www.internetsociety.org/deploy360/blog/2014/05/goodbye-ipv4-iana-startsallocating-final-address-blocks/. This means the IPv4 address space is close to being fully
allocated. As a result, it will soon become difficult to allocate more IPv4 addresses to an
application that has experienced growth, or is expected to scale out, due to the lack of
unallocated IPv4 address blocks.
For network focused applications the future is the IPv6 protocol. IPv6 increases the address
space significantly, fixes long standing issues in the IPv4 protocol, and will become an
essential for network focused applications in the future.
Neutron supports IPv6 when configured to take advantage of the feature. To enable it,
simply create an IPv6 subnet in OpenStack Neutron and use IPv6 prefixes when creating
security groups.

Asymmetric Links

When designing a network architecture, the traffic patterns of an application will heavily
influence the allocation of total bandwidth and the number of links that are used to send
and receive traffic. Applications that provide file storage for customers will allocate
bandwith and links to favor incoming traffic, whereas video streaming applications will
allocate bandwidth and links to favor outgoing traffic.

Performance

It is important to analyze the applications' tolerance for latency and jitter when designing an
environment to support network focused applications. Certain applications, for example
VoIP, are less tolerant of latency and jitter. Where latency and jitter are concerned, certain
applications may require tuning of QoS parameters and network device queues to ensure
that they are queued for transmit immediately or guaranteed minimum bandwidth. Since
OpenStack currently does not support these functions, some considerations may need to be
made for the network plugin selected.

The location of a service may also impact the application or consumer experience. If an
application is designed to serve differing content to differing users it will need to be designed
to properly direct connections to those specific locations. Use a multi-site installation for
these situations, where appropriate.
OpenStack networking can be implemented in two separate ways. The legacy nova-network
provides a flat DHCP network with a single broadcast domain. This implementation does not
support tenant isolation networks or advanced plugins, but it is currently the only way to
implement a distributed layer 3 agent using the multi_host configuration. Neutron is the
official current implementation of OpenStack Networking. It provides a pluggable
architecture that supports a large variety of network methods. Some of these include a layer
2 only provider network model, external device plugins, or even OpenFlow controllers.
Networking at large scales becomes a set of boundary questions. The determination of how
large a layer 2 domain needs to be is based on the amount of nodes within the domain and
the amount of broadcast traffic that passes between instances. Breaking layer 2 boundaries
may require the implementation of overlay networks and tunnels. This decision is a
balancing act between the need for a smaller overhead or a need for a smaller domain.
When selecting network devices, be aware that making this decision based on largest port
density often comes with a drawback. Aggregation switches and routers have not all kept
pace with Top of Rack switches and may induce bottlenecks on north-south traffic. As a
result, it may be possible for massive amounts of downstream network utilization to impact
upstream network devices, impacting service to the cloud. Since OpenStack does not
currently provide a mechanism for traffic shaping or rate limiting, it is necessary to
implement these features at the network hardware level.

Operational Considerations

Network focused OpenStack clouds have a number of operational considerations that will
influence the selected design. Topics including, but not limited to, dynamic routing of static
routes, service level agreements, and ownership of user management all need to be
considered.
One of the first required decisions is the selection of a telecom company or transit provider.
This is especially true if the network requirements include external or site-to-site network
connectivity.
Additional design decisions need to be made about monitoring and alarming. These can be
an internal responsibility or the responsibility of the external provider. In the case of using
an external provider, SLAs will likely apply. In addition, other operational considerations such
as bandwidth, latency, and jitter can be part of a service level agreement.
The ability to upgrade the infrastructure is another subject for consideration. As demand for
network resources increase, operators will be required to add additional IP address blocks
and add additional bandwidth capacity. Managing hardware and software lifecycle events,
for example upgrades, decommissioning, and outages while avoiding service interruptions
for tenants, will also need to be considered.
Maintainability will also need to be factored into the overall network design. This includes
the ability to manage and maintain IP addresses as well as the use of overlay identifiers
including VLAN tag IDs, GRE tunnel IDs, and MPLS tags. As an example, if all of the IP
addresses have to be changed on a network, a process known as renumbering, then the
design needs to support the ability to do so.
Network focused applications themselves need to be addressed when concerning certain
operational realities. For example, the impending exhaustion of IPv4 addresses, the
migration to IPv6 and the utilization of private networks to segregate different types of
traffic that an application receives or generates. In the case of IPv4 to IPv6 migrations,
applications should follow best practices for storing IP addresses. It is further recommended
to avoid relying on IPv4 features that were not carried over to the IPv6 protocol or have
differences in implementation.
When using private networks to segregate traffic, applications should create private tenant
networks for database and data storage network traffic, and utilize public networks for
client-facing traffic. By segregating this traffic, quality of service and security decisions can

be made to ensure that each network has the correct level of service that it requires.
Finally, decisions must be made about the routing of network traffic. For some applications,
a more complex policy framework for routing must be developed. The economic cost of
transmitting traffic over expensive links versus cheaper links, in addition to bandwidth,
latency, and jitter requirements, can be used to create a routing policy that will satisfy
business requirements.
How to respond to network events must also be taken into consideration. As
an example, how load is transferred from one link to another during a failure scenario could
be a factor in the design. If network capacity is not planned correctly, failover traffic could
overwhelm other ports or network links and create a cascading failure scenario. In this
case, traffic that fails over to one link overwhelms that link and then moves to the
subsequent links until the all network traffic stops.

Architecture

Network focused OpenStack architectures have many similarities to other OpenStack
architecture use cases. There a number of very specific considerations to keep in mind when
designing for a network-centric or network-heavy application environment.

Inter-dependencies

Networks exist to serve a as medium of transporting data between systems. It is inevitable
that an OpenStack design have inter-dependencies with non-network portions of
OpenStack as well as on external systems. Depending on the specific workload, there may
be major interactions with storage systems both within and external to the OpenStack
environment. For example, if the workload is a content delivery network, then the
interactions with storage will be two-fold. There will be traffic flowing to and from the
storage array for ingesting and serving content in a north-south direction. In addition, there
is replication traffic flowing in an east-west direction.
Compute-heavy workloads may also induce interactions with the network. Some high
performance compute applications require network-based memory mapping and data
sharing and, as a result, will induce a higher network load when they transfer results and
data sets. Others may be highly transactional and issue transaction locks, perform their
functions and rescind transaction locks at very high rates. This also has an impact on the
network performance.
Some network dependencies are going to be external to OpenStack. While Neutron is
capable of providing network ports, IP addresses, some level of routing, and overlay
networks, there are some other functions that it cannot provide. For many of these, external
systems or equipment may be required to fill in the functional gaps. Hardware load
balancers are an example of equipment that may be necessary to distribute workloads or
offload certain functions. Note that, as of the icehouse release, dynamic routing is
currently in its infancy within OpenStack and may need to be implemented either by an
external device or a specialized service instance within OpenStack. Tunneling is a feature
provided by Neutron, however it is constrained to a Neutron-managed region. If the need
arises to extend a tunnel beyond the OpenStack region to either another region or an
external system, it is necessary to implement the tunnel itself outside OpenStack or by using
a tunnel management system to map the tunnel or overlay to an external tunnel. OpenStack
does not currently provide quotas for network resources. Where network quotas are
required, it is necessary to implement quality of service management outside of OpenStack.
In many of these instances, similar solutions for traffic shaping or other network functions

will be needed.
Depending on the selected design, Neutron itself may not even support the required layer 3
network functionality. If it is necessary or advantageous to use the provider networking
mode of Neutron without running the layer 3 agent, then an external router will be required
to provide layer 3 connectivity to outside systems.
Interaction with orchestration services is inevitable in larger-scale deployments. Heat is
capable of allocating network resource defined in templates to map to tenant networks and
for port creation, as well as allocating floating IPs. If there is a requirement to define and
manage network resources in using orchestration, it is recommended that the design include
OpenStack Orchestration to meet the demands of users.

Design Impacts

A wide variety of factors can affect a network focused OpenStack architecture. While there
are some considerations shared with a general use case, specific workloads related to
network requirements will influence network design decisions.
One decision includes whether or not to use Network Address Translation (NAT) and where
to implement it. If there is a requirement for floating IPs to be available instead of using
public fixed addresses then NAT is required. This can be seen in network management
applications that rely on an IP endpoint. An example of this is a DHCP relay that needs to
know the IP of the actual DHCP server. In these cases it is easier to automate the
infrastructure to apply the target IP to a new instance rather than reconfigure legacy or
external systems for each new instance.
NAT for floating IPs managed by Neutron will reside within the hypervisor but there are also
versions of NAT that may be running elsewhere. If there is a shortage of IPv4
addresses there are two common methods to mitigate this externally to OpenStack. The
first is to run a load balancer either within OpenStack as a instance, or use an external load
balancing solution. In the internal scenario, load balancing software, such as HAproxy, can
be managed with Neutron's Load Balancer as a Service (LBaaS). This is specifically to
manage the Virtual IP (VIPs) while a dual-homed connection from the HAproxy instance
connects the public network with the tenant private network that hosts all of the content
servers. In the external scenario, a load balancer would need to serve the VIP and also be
joined to the tenant overlay network through external means or routed to it via private
addresses.

Another kind of NAT that may be useful is protocol NAT. In some cases it may be desirable
to use only IPv6 addresses on instances and operate either an instance or an external
service to provide a NAT-based transition technology such as NAT64 and DNS64. This
provides the ability to have a globally routable IPv6 address while only consuming IPv4
addresses as necessary or in a shared manner.
Application workloads will affect the design of the underlying network architecture. If a
workload requires network-level redundancy, the routing and switching architecture
will have to accommodate this. There are differing methods for providing this that are
dependent on the network hardware selected, the performance of the hardware, and which
networking model is deployed. Some examples of this are the use of Link aggregation (LAG)
or Hot Standby Router Protocol (HSRP). There are also the considerations of whether to
deploy Neutron or Nova-network and which plugin to select for Neutron. If using an
external system, Neutron will need to be configured to run layer 2 with a provider network
configuration. For example, it may be necessary to implement HSRP to terminate layer 3
connectivity.
Depending on the workload, overlay networks may or may not be a recommended
configuration. Where application network connections are small, short lived or bursty,
running a dynamic overlay can generate as much bandwidth as the packets it carries. It also
can induce enough latency to cause issues with certain applications. There is an impact to
the device generating the overlay which, in most installations, will be the hypervisor. This
will cause performance degradation on packet per second and connection per second rates.
Overlays also come with a secondary option that may or may not be appropriate to
a specific workload. While all of them will operate in full mesh by default, there might be
good reasons to disable this function because it may cause excessive overhead for some
workloads. Conversely, other workloads will operate without issue. For example, most web
services applications will not have major issues with a full mesh overlay network, while
some network monitoring tools or storage replication workloads will have performance
issues with throughput or excessive broadcast traffic.
A design decision that many overlook is a choice of layer 3 protocols. While OpenStack was
initially built with only IPv4 support, Neutron now supports IPv6 and dual-stacked
networks. Note that, as of the icehouse release, this only includes stateless address
autoconfiguration but the work is in progress to support stateless and stateful dhcpv6 as
well as IPv6 floating IPs without NAT. Some workloads become possible through the use of
IPv6 and IPv6 to IPv4 reverse transition mechanisms such as NAT64 and DNS64 or 6to4,

because these options are available. This will alter the requirements for any address plan as
single-stacked and transitional IPv6 deployments can alleviate the need for IPv4 addresses.
As of the icehouse release, OpenStack has limited support for dynamic routing, however
there are a number of options available by incorporating third party solutions to implement
routing within the cloud including network equipment, hardware nodes, and
instances. Some workloads will perform well with nothing more than static routes and
default gateways configured at the layer 3 termination point. In most cases this will suffice,
however some cases require the addition of at least one type of dynamic routing protocol if
not multiple protocols. Having a form of interior gateway protocol (IGP) available to the
instances inside an OpenStack installation opens up the possibility of use cases for anycast
route injection for services that need to use it as a geographic location or failover
mechanism. Other applications may wish to directly participate in a routing protocol, either
as a passive observer as in the case of a looking glass, or as an active participant in the form
of a route reflector. Since an instance might have a large amount of compute and memory
resources, it is trivial to hold an entire unpartitioned routing table and use it to provide
services such as network path visibility to other applications or as a monitoring tool.
A lesser known, but harder to diagnose issue, is that of path Maximum Transmission
Unit (MTU) failures. It is less of an optional design consideration and more a design warning
that MTU must be at least large enough to handle normal traffic, plus any overhead from an
overlay network, and the desired layer 3 protocol. Adding externally built tunnels will further
lessen the MTU packet size making it imperative to pay attention to the fully calculated MTU
as some systems may be configured to ignore or drop path MTU discovery packets.

Tunables

Networking configurables related to an OpenStack architecture design should be considered
when designing for network intensive workloads include MTU and QoS. Some workloads
will require a larger MTU than normal based on a requirement to transfer large blocks of
data. When providing network service for applications such as video streaming or storage
replication, it is recommended to ensure that both OpenStack hardware nodes and the
supporting network equipment are configured for jumbo frames where possible. This will
allow for a better utilization of available bandwidth. Configuration of jumbo frames should
be done across the complete path the packets will traverse. If one network component is
not capable of handling jumbo frames then the entire path will revert to the default MTU.
Quality of Service (QoS) also has a great impact on network intensive workloads by providing

instant service to packets which have a higher priority due to their ability to be impacted by
poor network performance. In applications such as Voice over IP (VoIP) differentiated
services code points are a near requirement for proper operation. QoS can also be used in
the opposite direction for mixed workloads to prevent low priority but high bandwidth
applications, for example backup services, video conferencing or file sharing, from blocking
bandwidth that is needed for the proper operation of other workloads. It is possible to tag
file storage traffic as a lower class, such as best effort or scavenger, to allow the higher
priority traffic through. In cases where regions within a cloud might be geographically
distributed it may also be necessary to plan accordingly to implement WAN optimization to
combat latency or packet loss.

Prescriptive Examples
Web Service Platform

A large-scale web application has been designed with cloud principles in mind. The
application is designed to scale horizontally in a bursting fashion and will generate a
high instance count. The application requires an SSL connection to secure data and must not
lose connection state to individual servers.
An example design for this workload is depicted in the figure below. In this example, a
hardware load balancer is configured to provide SSL offload functionality and to connect to
tenant networks in order to reduce address consumption. This load balancer is linked to the
routing architecture as it will service the VIP for the application. The router and load balancer
are configured with GRE tunnel ID of the application's tenant network and provided an IP
address within the tenant subnet but outside of the address pool. This is to ensure that the
load balancer can communicate with the application's HTTP servers without requiring the
consumption of a public IP address.
Since sessions must remain until closing, the routing and switching architecture is designed
for high availability. Switches are meshed to each hypervisor and to each other, and also
provide an MLAG implementation to ensure layer 2 connectivity does not fail. Routers are
configured with VRRP and fully meshed with switches to ensure layer 3 connectivity. Since
GRE is used as an overlay network, Neutron is installed and configured to use the Open
vSwitch agent in GRE tunnel mode. This ensures all devices can reach all other devices and
that tenant networks can be created for private addressing links to the load balancer.

Components

A web service architecture has many options and optional components. Due to this, it can fit
into a large number of other OpenStack designs however a few key components will need to
be in place to handle the nature of most webscale workloads. The user needs the
following components:
OpenStack Controller services (Image, Identity, Networking and supporting services such
as MariaDB and RabbitMQ)
OpenStack Compute running KVM hypervisor
OpenStack Object Storage
OpenStack Orchestration
OpenStack Telemetry
Beyond the normal Keystone, Nova, Glance and Swift components, Heat is a recommended

component to handle properly scaling the workloads to adjust to demand. Ceilometer will
also need to be included in the design due to the requirement for auto-scaling. Web services
tend to be bursty in load, have very defined peak and valley usage patterns and, as a
result, benefit from automatic scaling of instances based upon traffic. At a network level, a
split network configuration will work well with databases residing on private tenant
networks since these do not emit a large quantity of broadcast traffic and may need to
interconnect to some databases for content.

Load Balancing

Load balancing was included in this design to spread requests across multiple instances.
This workload scales well horizontally across large numbers of instances. This allows
instances to run without publicly routed IP addresses and simply rely on the load balancer
for the service to be globally reachable. Many of these services do not require direct server
return. This aids in address planning and utilization at scale since only the virtual IP (VIP)
must be public.

Overlay Networks

OpenStack Networking using the Open vSwitch GRE tunnel mode was included in the design
to provide overlay functionality. In this case, the layer 3 external routers will be in a pair with
VRRP and switches should be paired with an implementation of MLAG running to ensure
that there is no loss of connectivity with the upstream routing infrastructure.

Performance Tuning

Network level tuning for this workload is minimal. Quality-of-Service (QoS) will be applied to
these workloads for a middle ground Class Selector depending on existing policies. It will be
higher than a best effort queue but lower than an Expedited Forwarding or Assured
Forwarding queue. Since this type of application generates larger packets with longer-lived
connections, bandwidth utilization can be optimized for long duration TCP. Normal
bandwidth planning applies here with regards to benchmarking a session's usage multiplied
by the expected number of concurrent sessions with overhead.

Network Functions

Network functions is a broad category but encompasses workloads that support the rest of
a system's network. These workloads tend to consist of large amounts of small packets that
are very short lived, such as DNS queries or SNMP traps. These messages need to arrive
quickly and do not deal with packet loss as there can be a very large volume of them. There
are a few extra considerations to take into account for this type of workload and this can
change a configuration all the way to the hypervisor level. For an application that generates

10 TCP sessions per user with an average bandwidth of 512 kilobytes per second per flow
and expected user count of ten thousand concurrent users, the expected bandwidth plan is
approximately 4.88 gigabits per second.
The supporting network for this type of configuration needs to have a low latency and
evenly distributed availability. This workload benefits from having services local to the
consumers of the service. A multi-site approach is used as well as deploying many copies of
the application to handle load as close as possible to consumers. Since these applications
function independently, they do not warrant running overlays to interconnect tenant
networks. Overlays also have the drawback of performing poorly with rapid flow setup and
may incur too much overhead with large quantities of small packets and are therefore not
recommended.
QoS is desired for some workloads to ensure delivery. DNS has a major impact on the load
times of other services and needs to be reliable and provide rapid responses. It is to
configure rules in upstream devices to apply a higher Class Selector to DNS to ensure faster
delivery or a better spot in queuing algorithms.

Cloud Storage

Another common use case for OpenStack environments is to provide a cloud based file
storage and sharing service. While this may initially be considered to be a storage focused
use case there are also major requirements on the network side that place it in the realm of
requiring a network focused architecture. An example for this application is cloud backup.
There are two specific behaviors of this workload that have major and different impacts on
the network. Since this is both an externally facing service and internally replicating
application there are both North-South and East-West traffic considerations.
North-South traffic is primarily user facing. This means that when a user uploads content
for storage it will be coming into the OpenStack installation. Users who download this
content will be drawing traffic from the OpenStack installation. Since the service is intended
primarily as a backup the majority of the traffic will be southbound into the environment. In
this case it is beneficial to configure a network to be asymmetric downstream as the traffic
entering the OpenStack installation will be greater than traffic leaving.
East-West traffic is likely to be fully symmetric. Since replication will originate from any
node and may target multiple other nodes algorithmically, it is less likely for this traffic to
have a larger volume in any specific direction. However this traffic may interfere with north-

south traffic.

This application will prioritize the North-South traffic over East-West traffic as it is the
customer-facing data. QoS is implemented on East-West traffic to be a lower priority Class
Selector, while North-South traffic requires a higher level in the priority queue because of
this.
The network design in this case is less dependant on availability and more dependant on
being able to handle high bandwidth. As a direct result, it is beneficial to forego redundant
links in favor of bonding those connections. This increases available bandwidth. It is also
beneficial to configure all devices in the path, including OpenStack, to generate and pass
jumbo frames.

MULTI-SITE

Introduction

A multi-site OpenStack environment is one in which services located in more than one data
center are used to provide the overall solution. Usage requirements of different multi-site
clouds may vary widely, however they share some common needs. OpenStack is capable of
running in a multi-region configuration allowing some parts of OpenStack to effectively
manage a grouping of sites as a single cloud. With some careful planning in the design
phase, OpenStack can act as an excellent multi-site cloud solution for a multitude of needs.
Some use cases that might indicate a need for a multi-site deployment of OpenStack
include:
An organization with a diverse geographic footprint.
Geo-location sensitive data.
Data locality, in which specific data or functionality should be close to users.

User Requirements

A multi-site architecture is complex and has its own risks and considerations, therefore it is
important to make sure when contemplating the design such an architecture that it meets
the user and business requirements.

Legal Requirements

Many jurisdictions have legislative and regulatory requirements governing the storage and
management of data in cloud environments. Common areas of regulation include:
Data retention policies ensuring storage of persistent data and records management
to meet data archival requirements.
Data ownership policies governing the possession and responsibility for data.
Data sovereignty policies governing the storage of data in foreign countries or
otherwise separate jurisdictions.
Data compliance policies governing types of information that needs to reside in
certain locations due to regular issues and, more importantly, cannot reside in other
locations for the same reason.
Examples of such legal frameworks include the data protection framework of the European
Union (http://ec.europa.eu/justice/data-protection) and the requirements of the Financial
Industry Regulatory Authority (http://www.finra.org/Industry/Regulation/FINRARules) in
the United States. Consult a local regulatory body for more information.

Workload Characteristics

The expected workload is a critical requirement that needs to be captured to guide decisionmaking. An understanding of the workloads in the context of the desired multi-site
environment and use case is important. Another way of thinking about a workload is to think
of it as the way the systems are used. A workload could be a single application or a suite of
applications that work together. It could also be a duplicate set of applications that need to
run in multiple cloud environments. Often in a multi-site deployment the same workload
will need to work identically in more than one physical location.
This multi-site scenario likely includes one or more of the other scenarios in this book with
the additional requirement of having the workloads in two or more locations. The following
are some possible scenarios:

User Proximity of the workload (depending on the region)

For many use cases the proximity of the user to their workloads has a direct influence on the

performance of the application and therefore should be taken into consideration in the
design. Certain applications require zero to minimal latency that can only be achieved by
deploying the cloud in multiple locations. These locations could be in different data centers,
cities, countries or geographical regions, depending on the user requirement and location of
the users.

Consistency of images and templates across different sites

It is essential that the deployment of instances is consistent across the different sites. This
needs to be built into the infrastructure. If OpenStack Object Store is used as a backend for
Glance, it is possible to create repositories of consistent images across multiple sites. Having
a central endpoint with multiple storage nodes will allow for a consistent centralized storage
for each and every site.
Not using a centralized object store will increase operational overhead so that a consistent
image library can be maintained. This could include development of a replication mechanism
to handle the transport of images and the changes to the images across multiple sites.

High Availability

If high availability is a requirement to provide continuous infrastructure operations, a basic
requirement of High Availability should be defined.

OpenStack components

The OpenStack management components need to have a basic and minimal level of
redundancy. The simplest example is the loss of any single site has no significant impact on
the availability of the OpenStack services of the entire infrastructure.
The OpenStack High Availability Guide (http://docs.openstack.org/high-availabilityguide/content/) contains more information on how to provide redundancy for the
OpenStack components.

Network Connectivity

Multiple network links should be deployed between sites to provide redundancy for all
components. This includes storage replication, which should be isolated to a dedicated
network or VLAN with the ability to assign QoS to control the replication traffic or provide
priority for this traffic. Note that if the data store is highly changeable, the network
requirements could have a significant effect on the operational cost of maintaining the sites.

Storage redundancy

The ability to maintain object availability in both sites has significant implications on the
object storage design and implementation. It will also have a significant impact on the WAN
network design between the sites.

Three plus multi-site architectures

Connecting more than two sites increases the challenges and adds more complexity to the
design considerations. Multi-site implementations require extra planning to address the
additional topology complexity used for internal and external connectivity. Some options
include full mesh topology, hub spoke, spine leaf, or 3d Torus.

Cloud aware applications

Not all the applications running in a cloud are cloud-aware. If that is the case, there should
be clear measures and expectations to define what the infrastructure can support and, more
importantly, what it cannot. An example would be shared storage between sites. It is
possible, however such a solution is not native to OpenStack and requires a third-party
hardware vendor to fulfill such a requirement. Another example can be seen in applications
that are able to consume resources in object storage directly. These applications need to be
cloud aware to make good use of an OpenStack Object Store.

Risk Profile
Application readiness

Some applications are tolerant of the lack of synchronized object storage, while others may
need those objects to be replicated and available across regions. Understanding of how the
cloud implementation impacts new and existing applications is important for risk mitigation
and the overall success of a cloud project. Applications may have to be written to expect an
infrastructure with little to no redundancy. Existing applications not developed with the
cloud in mind may need to be rewritten.

Cost

The requirement of having more than one site has a cost attached to it. The greater the
number of sites, the greater the cost and complexity. Costs can be broken down into the
following categories
Compute Resources
Networking resources
Replication
Storage
Management

Operational costs

Site Loss and Recovery

Outages can cause loss of partial or full functionality of a site. Strategies should be
implemented to understand and plan for recovery scenarios.
The deployed applications need to continue to function and, more importantly,
consideration should be taken of the impact on the performance and reliability of the
application when a site is unavailable.
It is important to understand what will happen to replication of objects and data
between the sites when a site goes down. If this causes queues to start building up,
considering how long these queues can safely exist until something explodes.
Ensure determination of the method for resuming proper operations of a site when it
comes back online after a disaster. It is recommended to architect the recovery to avoid
race conditions.

Security
Compliance and Geo-location

An organization could have certain legal obligations and regulatory compliance
measures which could require certain workloads or data to not be located in certain regions.

Auditing

A well thought-out auditing strategy is important in order to be able to quickly track down
issues. Keeping track of changes made to security groups and tenant changes can be useful
in rolling back the changes if they affect production. For example, if all security group rules
for a tenant disappeared, the ability to quickly track down the issue would be important for
operational and legal reasons.

Separation of duties

A common requirement is to define different roles for the different cloud administration
functions. An example would be a requirement to segregate the duties and permissions by
site.

Authentication between sites

Ideally it is best to have a single authentication domain and not need a separate
implementation for each and every site. This will, of course, require an authentication
mechanism that is highly available and distributed to ensure continuous operation.
Authentication server locality is also something that might be needed as well and should be
planned for.

Technical Considerations

There are many technical considerations to take into account with regard to designing a
multi-site OpenStack implementation. An OpenStack cloud can be designed in a variety of
ways to handle individual application needs. A multi-site deployment will have additional
challenges compared to single site installations and will therefore be a more complex
solution.

Capacity

When determining capacity options be sure to take into account not just the technical
issues, but also the economic or operational issues that might arise from specific decisions.
Inter-site link capacity describes the capabilities of the connectivity between the
different OpenStack sites. This includes parameters such as bandwidth, latency, whether or
not a link is dedicated, and any business policies applied to the connection. The capability
and number of the links between sites will determine what kind of options may be available
for deployment. For example, if two sites have a pair of high-bandwidth links available
between them, it may be wise to configure a separate storage replication network between
the two sites to support a single Swift endpoint and a shared object storage capability
between them. (An example of this technique, as well as a configuration walkthrough, is
available at
http://docs.openstack.org/developer/swift/replication_network.html#dedicatedreplication-network). Another option in this scenario is to build a dedicated set of tenant
private networks across the secondary link using overlay networks with a third party
mapping the site overlays to each other.
The capacity requirements of the links between sites will be driven by application behavior.
If the latency of the links is too high, certain applications that use a large number of small
packets, for example RPC calls, may encounter issues communicating with each other or
operating properly. Additionally, OpenStack may encounter similar types of issues.
To mitigate this, tuning of the Keystone call timeouts may be necessary to prevent issues
authenticating against a central Identity Service.
Another capacity consideration when it comes to networking for a multi-site deployment is
the available amount and performance of overlay networks for tenant networks. If using
shared tenant networks across zones, it is imperative that an external overlay manager or
controller be used to map these overlays together. It is necessary to ensure the amount of
possible IDs between the zones are identical. Note that, as of the Icehouse release, Neutron

was not capable of managing tunnel IDs across installations. This means that if one site runs
out of IDs, but other does not, that tenant's network will be unable to reach the other site.
Capacity can take other forms as well. The ability for a region to grow depends on scaling
out the number of available compute nodes. This topic is covered in greater detail in the
section for compute-focused deployments. However, it should be noted that cells may be
necessary to grow an individual region beyond a certain point. This point depends on the size
of your cluster and the ratio of virtual machines per hypervisor.
A third form of capacity comes in the multi-region-capable components of OpenStack.
Centralized Object Storage is capable of serving objects through a single namespace across
multiple regions. Since this works by accessing the object store via swift proxy, it is possible
to overload the proxies. There are two options available to mitigate this issue. The first is to
deploy a large number of swift proxies. The drawback to this is that the proxies are not loadbalanced and a large file request could continually hit the same proxy. The other way to
mitigate this is to front-end the proxies with a caching HTTP proxy and load balancer. Since
swift objects are returned to the requester via HTTP, this load balancer would alleviate the
load required on the swift proxies.

Utilization

While constructing a multi-site OpenStack environment is the goal of this guide, the real test
is whether an application can utilize it.
Identity is normally the first interface for the majority of OpenStack users. Interacting with
Keystone is required for almost all major operations within OpenStack. Therefore, it is
important to ensure that you provide users with a single URL for Keystone authentication.
Equally important is proper documentation and configuration of regions within Keystone.
Each of the sites defined in your installation is considered to be a region in Keystone
nomenclature. This is important for the users of the system, when reading Keystone
documentation, as it is required to define the Region name when providing actions to an API
endpoint or in Horizon.
Load balancing is another common issue with multi-site installations. While it is still possible
to run HAproxy instances with load balancer as a service, these will be local to a specific
region. Some applications may be able to cope with this via internal mechanisms. Others,
however, may require the implementation of an external system including global services
load balancers or anycast-advertised DNS.

Depending on the storage model chosen during site design, storage replication and
availability will also be a concern for end-users. If an application is capable of understanding
regions, then it is possible to keep the object storage system separated by region. In this
case, users who want to have an object available to more than one region will need to do the
cross-site replication themselves. With a centralized swift proxy, however, the user may
need to benchmark the replication timing of the Swift backend. Benchmarking allows the
operational staff to provide users with an understanding of the amount of time required for
a stored or modified object to become available to the entire environment.

Performance

Determining the performance of a multi-site installation involves considerations that do
not come into play in a single-site deployment. Being a distributed deployment, multi-site
deployments incur a few extra penalties to performance in certain situations.
Since multi-site systems can be geographically separated, they may have worse than
normal latency or jitter when communicating across regions. This can especially impact
systems like the OpenStack Identity service when making authentication attempts from
regions that do not contain the centralized Keystone implementation. It can also affect
certain applications which rely on remote procedure call (RPC) for normal operation. An
example of this can be seen in High Performance Computing workloads.
Storage availability can also be impacted by the architecture of a multi-site deployment.
A centralized Object Storage Service requires more time for an object to be available to
instances locally in regions where the object was not created. Some applications may need
to be tuned to account for this effect. Block storage does not currently have a method for
replicating data across multiple regions, so applications that depend on available block
storage will need to manually cope with this limitation by creating duplicate block storage
entries in each region.

Security

Securing a multi-site OpenStack installation also brings extra challenges. Tenants may
expect a tenant-created network to be secure. In a multi-site installation the use of a nonprivate connection between sites may be required. This may mean that traffic would be
visible to third parties and, in cases where an application requires security, this issue will
require mitigation. Installing a VPN or encrypted connection between sites is recommended
in such instances.

Another security consideration with regard to multi-site deployments is Identity.
Authentication in a multi-site deployment should be centralized. Centralization provides a
single authentication point for users across the deployment, as well as a single point of
administration for traditional create, read, update and delete operations. Centralized
authentication is also useful for auditing purposes because all authentication tokens
originate from the same source.
Just as tenants in a single-site deployment need isolation from each other, so do tenants in
multi-site installations. The extra challenges in multi-site designs revolve around ensuring
that tenant networks function across regions. Unfortunately, OpenStack Networking does
not presently support a mechanism to provide this functionality, therefore an external
system may be necessary to manage these mappings. Tenant networks may contain
sensitive information requiring that this mapping be accurate and consistent to ensure that
a tenant in one site does not connect to a different tenant in another site.

OpenStack Components

Most OpenStack installations require a bare minimum set of pieces to function. These
include Keystone for authentication, Nova for compute, Glance for image storage, Neutron
for networking, and potentially an object store in the form of Swift. Bringing multi-site into
play also demands extra components in order to coordinate between regions. Centralized
Keystone is necessary to provide the single authentication point. Centralized Horizon is also
recommended to provide a single login point and a mapped experience to the API and CLI
options available. If necessary, a centralized Swift may be used and will require the
installation of the Swift proxy service.
It may also be helpful to install a few extra options in order to facilitate certain use cases. For
instance, installing Designate may assist in automatically generating DNS domains for each
region with an automatically-populated zone full of resource records for each instance. This
facilitates using DNS as a mechanism for determining which region would be selected for
certain applications.
Another useful tool for managing a multi-site installation is Heat. Heat allows the use of
templates to define a set of instances to be launched together or for scaling existing sets. It
can also be used to setup matching or differentiated groupings based on regions. For
instance, if an application requires an equally balanced number of nodes across sites, the
same heat template can be used to cover each site with small alterations to only the region
name.

Operational Considerations
Installation
Tooling

Deployment of a multi-site OpenStack cloud using regions requires that the service catalog
contains per-region entries for each service deployed other than the Identity service itself.
There is limited support amongst currently available off-the-shelf OpenStack deployment
tools for defining multiple regions in this fashion.
Deployers must be aware of this and provide the appropriate customization of the service
catalog for their site either manually or via customization of the deployment tools in use.
Note that, as of the Icehouse release, documentation for implementing this feature is in
progress. See this bug for more information: https://bugs.launchpad.net/openstackmanuals/+bug/1340509

Licensing

Multi-site OpenStack deployments present additional licensing considerations over and
above regular OpenStack clouds, particularly where site licenses are in use to provide cost
efficient access to software licenses. The licensing for host operating systems, guest
operating systems, OpenStack distributions (if applicable), software-defined
infrastructure including network controllers and storage systems, and even individual
applications need to be evaluated in light of the multi-site nature of the cloud.
Topics to consider include:
The specific definition of what constitutes a site in the relevant licenses, as the term
does not necessarily denote a geographic or otherwise physically isolated location in the
traditional sense.
Differentiations between "hot" (active) and "cold" (inactive) sites where significant
savings may be made in situations where one site is a cold standby for disaster recovery
purposes only.
Certain locations might require local vendors to provide support and services for each
site provides challenges, but will vary on the licensing agreement in place.

Logging and Monitoring

Logging and monitoring does not significantly differ for a multi-site OpenStack cloud. The
same well known tools described in the Operations Guide
(http://docs.openstack.org/openstack-ops/content/logging_monitoring.html) remain

applicable. Logging and monitoring can be provided both on a per-site basis and in a
common centralized location.
When attempting to deploy logging and monitoring facilities to a centralized location, care
must be taken with regards to the load placed on the inter-site networking links.

Upgrades

In multi-site OpenStack clouds deployed using regions each site is, effectively, an
independent OpenStack installation which is linked to the others by using
centralized services such as Identity which are shared between sites. At a high level the
recommended order of operations to upgrade an individual OpenStack environment is
(http://docs.openstack.org/openstack-ops/content/ops_upgrades-general-steps.html):
1. Upgrade the OpenStack Identity Service (Keystone).
2. Upgrade the OpenStack Image Service (Glance).
3. Upgrade OpenStack Compute (Nova), including networking components.
4. Upgrade OpenStack Block Storage (Cinder).
5. Upgrade the OpenStack dashboard.(Horizon)
The process for upgrading a multi-site environment is not significantly different:
1. Upgrade the shared OpenStack Identity Service (Keystone) deployment.
2. Upgrade the OpenStack Image Service (glance) at each site.
3. Upgrade OpenStack Compute (Nova), including networking components, at each site.
4. Upgrade OpenStack Block Storage (Cinder) at each site.
5. Upgrade the OpenStack dashboard (Horizon), at each site - or in the single central
location if it is shared.
Note that, as of the OpenStack Icehouse release, compute upgrades within each site can
also be performed in a rolling fashion. Compute controller services (API, Scheduler, and
Conductor) can be upgraded prior to upgrading of individual compute nodes. This maximizes
the ability of operations staff to keep a site operational for users of compute services while

performing an upgrade.

Maintainability
Quota Management

To prevent system capacities from being exhausted without notification, OpenStack
provides operators with the ability to define quotas. Quotas are used to set operational
limits and are currently enforced at the tenant (or project) level rather than at the user level.
Quotas are defined on a per-region basis. Operators may wish to define identical quotas for
tenants in each region of the cloud to provide a consistent experience, or even create a
process for synchronizing allocated quotas across regions. It is important to note that only
the operational limits imposed by the quotas will be aligned consumption of quotas by users
will not be reflected between regions.
For example, given a cloud with two regions, if the operator grants a user a quota of 25
instances in each region then that user may launch a total of 50 instances spread across
both regions. They may not, however, launch more than 25 instances in any single region.
For more information on managing quotas refer to Chapter 9. Managing Projects and Users
(http://docs.openstack.org/openstack-ops/content/projects_users.html) of the
OpenStack Operators Guide.

Policy Management

OpenStack provides a default set of Role Based Access Control (RBAC) policies, defined in a
policy.json file, for each service. Operators edit these files to customize the policies for their
OpenStack installation. If the application of consistent RBAC policies across sites is
considered a requirement, then it is necessary to ensure proper synchronization of the
policy.json files to all installations.
This must be done using normal system administration tools such as rsync as no
functionality for synchronizing policies across regions is currently provided within
OpenStack.

Documentation

Users must be able to leverage cloud infrastructure and provision new resources in the
environment. It is important that user documentation is accessible by users of the cloud
infrastructure to ensure they are given sufficient information to help them leverage the
cloud. As an example, by default OpenStack will schedule instances on a compute node

automatically. However, when multiple regions are available, it is left to the end user to
decide in which region to schedule the new instance. Horizon will present the user with the
first region in your configuration. The API and CLI tools will not execute commands unless a
valid region is specified. It is therefore important to provide documentation to your users
describing the region layout as well as calling out that quotas are region-specific. If a user
reaches his or her quota in one region, OpenStack will not automatically build new instances
in another. Documenting specific examples will help users understand how to operate the
cloud, thereby reducing calls and tickets filed with the helpdesk.

Architecture
Logical Diagram

This graphic is a high level diagram of a multiple site OpenStack architecture. Each site is an
OpenStack cloud but it may be necessary for the sites to be architected on different
versions. For example, if the second site is intended to be a replacement for the first site,
they would be different. Another common design would be a private OpenStack cloud with
replicated site that would be used for high availability or disaster recovery. The most
important design decision is how to configure the storage. It can be configured as a single
shared pool or separate pools, depending on the user and technical requirements.

OpenStack Services Architecture

The OpenStack Identity service, which is used by all other OpenStack components for
authorization and the catalog of service endpoints, supports the concept of regions. A region
is a logical construct that can be used to group OpenStack services that are in close
proximity to one another. The concept of regions is flexible; it may can contain OpenStack
service endpoints located within a distinct geographic region, or regions. It may be smaller in
scope, where a region is a single rack within a data center or even a single blade chassis, with

multiple regions existing in adjacent racks in the same data center.
The majority of OpenStack components are designed to run within the context of a single
region. The OpenStack Compute service is designed to manage compute resources within a
region, with support for subdivisions of compute resources by using Availability Zones and
Cells. The OpenStack Networking service can be used to manage network resources in the
same broadcast domain or collection of switches that are linked. The OpenStack Block
Storage service controls storage resources within a region with all storage resources residing
on the same storage network. Like the OpenStack Compute service, the OpenStack Block
Storage Service also supports the Availability Zone construct,which can be used to subdivide
storage resources.
The OpenStack Dashboard, OpenStack Identity Service, and OpenStack Object Storage
services are components that can each be deployed centrally in order to serve multiple
regions.

Storage

With multiple OpenStack regions, having a single OpenStack Object Storage Service
endpoint that delivers shared file storage for all regions is desirable. The Object Storage
service internally replicates files to multiple nodes. The advantages of this are that, if a file
placed into the Object Storage service is visible to all regions, it can be used by applications
or workloads in any or all of the regions. This simplifies high availability failover and disaster
recovery rollback.
In order to scale the Object Storage service to meet the workload of multiple regions,
multiple proxy workers are run and load-balanced, storage nodes are installed in each
region, and the entire Object Storage Service can be fronted by an HTTP caching layer. This is
done so client requests for objects can be served out of caches rather than directly from the
storage modules themselves, reducing the actual load on the storage network. In addition to
an HTTP caching layer, a caching layer like Memcache will be used to also cache objects
between the proxy and storage nodes.
If the cloud is designed without a single Object Storage Service endpoint for multiple regions,
and instead a separate Object Storage Service endpoint is made available in each region,
applications are required to handle synchronization (if desired) and other management
operations to ensure consistency across the nodes. For some applications, having multiple
Object Storage Service endpoints located in the same region as the application may be
desirable due to reduced latency, cross region bandwidth, and ease of deployment.

For the Block Storage service, the most important decisions are the selection of the storage
technology and whether or not a dedicated network is used to carry storage traffic from the
storage service to the compute nodes.

Networking

When connecting multiple regions together there are several design considerations. The
overlay network technology choice determines how packets are transmitted between
regions and how the logical network and addresses present to the application. If there are
security or regulatory requirements, encryption should be implemented to secure the traffic
between regions. For networking inside a region, the overlay network technology for tenant
networks is equally important. The overlay technology and the network traffic of an
application generates or receives can be either complementary or be at cross purpose. For
example, using an overlay technology for an application that transmits a large amount of
small packets could add excessive latency or overhead to each packet if not configured
properly.

Dependencies

The architecture for a multi-site installation of OpenStack is dependent on a number of
factors. One major dependency to consider is storage. When designing the storage system,
the storage mechanism needs to be determined. Once the storage type is determined, how it
will be accessed is critical. For example, it is recommended that storage should utilize a
dedicated network. Another concern is how the storage is configured to protect the data.
For example, the recovery point objective (RPO) and the recovery time objective (RTO). How
quickly can the recovery from a fault be completed, will determine how often the replication
of data be required. Ensure that enough storage is allocated to support the data protection
strategy.
Networking decisions include the encapsulation mechanism that will be used for the tenant
networks, how large the broadcast domains should be, and the contracted SLAs for the
interconnects.

Prescriptive Examples

Based on the needs of the intended workloads, there are multiple ways to build a multi-site
OpenStack installation. Below are example architectures based on different requirements.
These examples are meant as a reference, and not a hard and fast rule for deployments. Use
the previous sections of this chapter to assist in selecting specific components and
implementations based on specific needs.

Customer edge data centers

A large content provider needs to deliver content to customers that are geographically
dispersed. The workload is very sensitive to latency and needs a rapid response to endusers. After reviewing the user, technical and operational considerations, it is determined
beneficial to build a number of regions local to the customer’s edge. In this case rather than
build a few large, centralized data centers, the intent of the architecture is to provide a pair
of small data centers in locations that are closer to the customer. In this use case, spreading
applications out allows for different horizontal scaling than a traditional compute workload
scale. The intent is to scale by creating more copies of the application in closer proximity to
the users that need it most, in order to ensure faster response time to user requests. This
provider will deploy two datacenters at each of the four chosen regions. The implications of
this design are based around the method of placing copies of resources in each of the
remote regions. Swift objects, Glance images, and block storage will need to be manually
replicated into each region. This may be beneficial for some systems, such as the case of
content service, where only some of the content needs to exist in some but not all regions. A
centralized Keystone is recommended to ensure authentication and that access to the API
endpoints is easily manageable.
Installation of an automated DNS system such as Designate is highly recommended. Unless
an external Dynamic DNS system is available, application administrators will need a way to
manage the mapping of which application copy exists in each region and how to reach it.
Designate will assist by making the process automatic and by populating the records in the
each region's zone.
Telemetry for each region is also deployed, as each region may grow differently or be used
at a different rate. Ceilometer will run to collect each region's metrics from each of the
controllers and report them back to a central location. This is useful both to the end user
and the administrator of the OpenStack environment. The end user will find this method
useful, in that it is possible to determine if certain locations are experiencing higher load than
others, and take appropriate action. Administrators will also benefit by possibly being able

to forecast growth per region, rather than expanding the capacity of all regions
simultaneously, therefore maximizing the cost-effectiveness of the multi-site design.
One of the key decisions of running this sort of infrastructure is whether or not to provide a
redundancy model. Two types of redundancy and high availability models in this
configuration will be implemented. The first type revolves around the availability of the
central OpenStack components. Keystone will be made highly available in three central data
centers that will host the centralized OpenStack components. This prevents a loss of any
one of the regions causing an outage in service. It also has the added benefit of being able to
run a central storage repository as a primary cache for distributing content to each of the
regions.
The second redundancy topic is that of the edge data center itself. A second datacenter in
each of the edge regional locations will house a second region near the first. This ensures
that the application will not suffer degraded performance in terms of latency and
availability.
This figure depicts the solution designed to have both a centralized set of core data centers
for OpenStack services and paired edge data centers:

Geo-redundant load balancing

A large-scale web application has been designed with cloud principles in mind. The
application is designed provide service to application store, on a 24/7 basis. The company
has typical 2-tier architecture with a web front-end servicing the customer requests and a
NoSQL database backend storing the information.
As of late there has been several outages in number of major public cloud providers usually due to the fact these applications were running out of a single geographical location.
The design therefore should mitigate the chance of a single site causing an outage for their
business.
The solution would consist of the following OpenStack components:
A firewall, switches and load balancers on the public facing network connections.
OpenStack Controller services running, Networking, Horizon, Cinder and Nova
compute running locally in each of the three regions. The other services, Keystone, Heat
Ceilometer, Glance and Swift will be installed centrally - with nodes in each of the region
providing a redundant OpenStack Controller plane throughout the globe.

OpenStack Compute nodes running the KVM hypervisor.
OpenStack Object Storage for serving static objects such as images will be used to
ensure that all images are standardized across all the regions, and replicated on a regular
basis.
A Distributed DNS service available to all regions - that allows for dynamic update of
DNS records of deployed instances.
A geo-redundant load balancing service will be used to service the requests from the
customers based on their origin.
An autoscaling heat template will used to deploy the application in the three regions. This
template will include:
Web Servers, running Apache.
Appropriate user_data to populate the central DNS servers upon instance launch.
Appropriate Ceilometer alarms that maintain state of the application and allow for
handling of region or instance failure.
Another autoscaling Heat template will be used to deploy a distributed MongoDB shard
over the three locations - with the option of storing required data on a globally available
Swift container. according to the usage and load on the database server - additional shards
will be provisioned according to the thresholds defined in Ceilometer.
The reason that 3 regions were selected here was because of the fear of having abnormal
load on a single region in the event of a failure. Two datacenter would have been sufficient
had the requirements been met.
Heat is used because of the built-in functionality of autoscaling and auto healing in the
event of increased load. Additional configuration management tools, such as Puppet or Chef
could also have been used in this scenario, but were not chosen due to the fact that Heat
had the appropriate built-in hooks into the OpenStack cloud - whereas the other tools were
external and not native to OpenStack. In addition - since this deployment scenario was
relatively straight forward - the external tools were not needed.
Swift is used here to serve as a backend for Glance and Object storage since was the most
suitable solution for a globally distributed storage solution - with its own replication
mechanism. Home grown solutions could also have been used including the handling of
replication - but were not chosen, because Swift is already an intricate part of the
infrastructure - and proven solution.

An external load balancing service was used and not the LBaaS in Openstack because the
solution in Openstack is not redundant and does have any awareness of geo location.

Location-local service

A common use for a multi-site deployment of OpenStack, is for creating a Content Delivery
Network. An application that uses a location-local architecture will require low network
latency and proximity to the user, in order to provide an optimal user experience, in addition
to reducing the cost of bandwidth and transit, since the content resides on sites closer to the
customer, instead of a centralized content store that would require utilizing higher cost
cross country links.
This architecture usually includes a geolocation component that places user requests at the
closest possible node. In this scenario, 100% redundancy of content across every site is a
goal rather than a requirement, with the intent being to maximize the amount of content
available that is within a minimum number of network hops for any given end user. Despite
these differences, the storage replication configuration has significant overlap with that of a
geo-redundant load balancing use case.

In this example, the application utilizing this multi-site OpenStack install that is location
aware would launch web server or content serving instances on the compute cluster in each
site. Requests from clients will first be sent to a global services loadbalancer that
determines the location of the client, then routes the request to the closest OpenStack site
where the application completes the request.

HYBRID CLOUD

Introduction

Hybrid cloud, by definition, means that the design spans more than one cloud. An example
of this kind of architecture may include a situation in which the design involves more than
one OpenStack cloud (for example, an OpenStack-based private cloud and an OpenStackbased public cloud), or it may be a situation incorporating an OpenStack cloud and a nonOpenStack cloud (for example, an OpenStack-based private cloud that interacts with
Amazon Web Services). Bursting into an external cloud is the practice of creating new
instances to alleviate extra load where there is no available capacity in the private cloud.
Some situations that could involve hybrid cloud architecture include:
Bursting from a private cloud to a public cloud
Disaster recovery
Development and testing
Federated cloud, enabling users to choose resources from multiple providers
Hybrid clouds built to support legacy systems as they transition to cloud
As a hybrid cloud design deals with systems that are outside of the control of the cloud
architect or organization, a hybrid cloud architecture requires considering aspects of the
architecture that might not have otherwise been necessary. For example, the design may
need to deal with hardware, software, and APIs under the control of a separate organization.
Similarly, the degree to which the architecture is OpenStack-based will have an effect on the
cloud operator or cloud consumer's ability to accomplish tasks with native OpenStack tools.
By definition, this is a situation in which no single cloud can provide all of the necessary
functionality. In order to manage the entire system, users, operators and consumers will
need an overarching tool known as a cloud management platform (CMP). Any organization
that is working with multiple clouds already has a CMP, even if that CMP is the operator who
logs into an external web portal and launches a public cloud instance.
There are commercially available options, such as Rightscale, and open source options, such
as ManageIQ (http://manageiq.org/), but there is no single CMP that can address all needs
in all scenarios. Whereas most of the sections of this book talk about the aspects of
OpenStack, an architect needs to consider when designing an OpenStack architecture. This
section will also discuss the things the architect must address when choosing or building a
CMP to run a hybrid cloud design, even if the CMP will be a manually built solution.

User Requirements

Hybrid cloud architectures introduce additional complexities, particularly those that use
heterogeneous cloud platforms. As a result, it is important to make sure that design choices
match requirements in such a way that the benefits outweigh the inherent additional
complexity and risks.

Business considerations

Business considerations to make when designing a hybrid cloud deployment include:
Cost: A hybrid cloud architecture involves multiple vendors and technical
architectures. These architectures may be more expensive to deploy and maintain.
Operational costs can be higher because of the need for more sophisticated
orchestration and brokerage tools than in other architectures. In contrast, overall
operational costs might be lower by virtue of using a cloud brokerage tool to deploy the
workloads to the most cost effective platform.
Revenue opportunity: Revenue opportunities vary greatly based on the intent and use
case of the cloud. If it is being built as a commercial customer-facing product, consider
the drivers for building it over multiple platforms and whether the use of multiple
platforms make the design more attractive to target customers, thus enhancing the
revenue opportunity.
Time to Market: One of the most common reasons to use cloud platforms is to speed
the time to market of a new product or application. A business requirement to use
multiple cloud platforms may be because there is an existing investment in several
applications and it is faster to tie them together rather than migrating components and
refactoring to a single platform.
Business or technical diversity: Organizations already leveraging cloud-based services
may wish to embrace business diversity and utilize a hybrid cloud design to spread their
workloads across multiple cloud providers so that no application is hosted in a single
cloud provider.
Application momentum: A business with existing applications that are already in
production on multiple cloud environments may find that it is more cost effective to
integrate the applications on multiple cloud platforms rather than migrate them to a
single platform.

Legal Requirements

Many jurisdictions have legislative and regulatory requirements governing the storage and
management of data in cloud environments. Common areas of regulation include:

Data retention policies ensuring storage of persistent data and records management
to meet data archival requirements.
Data ownership policies governing the possession and responsibility for data.
Data sovereignty policies governing the storage of data in foreign countries or
otherwise separate jurisdictions.
Data compliance policies governing certain types of information needs to reside in
certain locations due to regular issues and, more importantly, cannot reside in other
locations for the same reason.
Examples of such legal frameworks include the data protection framework of the European
Union (http://ec.europa.eu/justice/data-protection/) and the requirements of the Financial
Industry Regulatory Authority (http://www.finra.org/Industry/Regulation/FINRARules/) in
the United States. Consult a local regulatory body for more information.

Workload Considerations

Defining what the word "workload" means in the context of a hybrid cloud environment is
important. Workload can be defined as the intended way the systems will be utilized, which
is often referred to as a “use case.” A workload can be a single application or a suite of
applications that work in concert. It can also be a duplicate set of applications that need to
run on multiple cloud environments.In a hybrid cloud deployment, the same workload will
often need to function equally well on radically different public and private cloud
environments. The architecture needs to address these potential conflicts, complexity, and
platform incompatibilities.
Some possible use cases for a hybrid cloud architecture include:
Dynamic resource expansion or "bursting": Another common reason to use a multiple
cloud architecture is a "bursty" application that needs additional resources at times. An
example of this case could be a retailer that needs additional resources during the
holiday selling season, but does not want to build expensive cloud resources to meet the
peak demand. They might have an OpenStack private cloud but want to burst to AWS or
some other public cloud for these peak load periods. These bursts could be for long or
short cycles ranging from hourly, monthly or yearly.
Disaster recovery-business continuity: The cheaper storage and instance
management makes a good case for using the cloud as a secondary site. The public cloud
is already heavily used for these purposes in combination with an OpenStack public or
private cloud.
Federated hypervisor-instance management: Adding self-service, chargeback and

transparent delivery of the right resources from a federated pool can be cost effective. In
a hybrid cloud environment, this is a particularly important consideration. Look for a
cloud that provides cross-platform hypervisor support and robust instance
management tools.
Application portfolio integration: An enterprise cloud delivers better application
portfolio management and more efficient deployment by leveraging self-service
features and rules for deployments based on types of use. A common driver for building
hybrid cloud architecture is to stitch together multiple existing cloud environments that
are already in production or development. In the interest of time to market, the
requirements may be to maintain the multiple clouds and just integrate the pieces
togethernot rationalize to one cloud environment, but instead to
Migration scenarios: A common reason to create a hybrid cloud architecture is to
allow the migration of applications between different clouds. This may be because the
application will be migrated permanently to a new platform, or it might be because the
application needs to be supported on multiple platforms going forward.
High availability: Another important reason for wanting a multiple cloud architecture
is to address the needs for high availability. By using a combination of multiple locations
and platforms, a design can achieve a level of availability that is not possible with a single
platform. This approach does add a significant amount of complexity.
In addition to thinking about how the workload will work on a single cloud, the design must
accommodate the added complexity of needing the workload to run on multiple cloud
platforms. The complexity of transferring workloads across clouds needs to be explored at
the application, instance, cloud platform, hypervisor, and network levels.

Tools Considerations

When working with designs spanning multiple clouds, the design must incorporate tools to
facilitate working across those multiple clouds. Some of the user requirements drive the
need for tools that will do the following functions:
Broker between clouds: Since the multiple cloud architecture assumes that there will
be at least two different and possibly incompatible platforms that are likely to have
different costs, brokering software is designed to evaluate relative costs between
different cloud platforms. These solutions are sometimes referred to as Cloud
Management Platforms (CMPs). Examples include Rightscale, Gravitent, Scalr,
CloudForms, and ManageIQ. These tools allow the designer to determine the right
location for the workload based on predetermined criteria.
Facilitate orchestration across the clouds: CMPs are tools are used to tie everything

together. Cloud orchestration tools are used to improve the management of IT
application portfolios as they migrate onto public, private, and hybrid cloud platforms.
These tools are an important consideration. Cloud orchestration tools are used for
managing a diverse portfolio of installed systems across multiple cloud platforms. The
typical enterprise IT application portfolio is still comprised of a few thousand
applications scattered over legacy hardware, virtualized infrastructure, and now dozens
of disjointed shadow public Infrastructure-as-a-Service (IaaS) and Software-as-aService (SaaS) providers and offerings.

Network Considerations

The network services functionality is an important factor to assess when choosing a CMP
and cloud provider. Considerations are functionality, security, scalability and HA. Verification
and ongoing testing of the critical features of the cloud endpoint used by the architecture
are important tasks.
Once the network functionality framework has been decided, a minimum functionality
test should be designed. This will ensure testing and functionality persists during and
after upgrades.
Scalability across multiple cloud providers may dictate which underlying network
framework you will choose in different cloud providers. It is important to have the
network API functions presented and to verify that functionality persists across all cloud
endpoints chosen.
High availability implementations vary in functionality and design. Examples of some
common methods are Active-Hot-Standby, Active-Passive and Active-Active. High
availability and a test framework needs to be developed to insure that the functionality
and limitations are well understood.
Security considerations include how data is secured between client and endpoint and
any traffic that traverses the multiple clouds, from eavesdropping to DoS activities.

Risk Mitigation and Management Considerations

Hybrid cloud architectures introduce additional risk because they add additional complexity
and potentially conflicting or incompatible components or tools. However, they also reduce
risk by spreading workloads over multiple providers. This means, if one was to go out of
business, the organization could remain operational.
Risks that will be heightened by using a hybrid cloud architecture include:
Provider availability or implementation details: This can range from the company
going out of business to the company changing how it delivers its services. Cloud

architectures are inherently designed to be flexible and changeable; paradoxically, the
cloud is both perceived to be rock solid and ever flexible at the same time.
Differing SLAs: Users of hybrid cloud environments potentially encounter some
losses through differences in service level agreements. A hybrid cloud design needs to
accommodate the different SLAs provided by the various clouds involved in the design,
and must address the actual enforceability of the providers' SLAs.
Security levels: Securing multiple cloud environments is more complex than securing
a single cloud environment. Concerns need to be addressed at, but not limited to, the
application, network, and cloud platform levels. One issue is that different cloud
platforms approach security differently, and a hybrid cloud design must address and
compensate for differences in security approaches. For example, AWS uses a relatively
simple model that relies on user privilege combined with firewalls.
Provider API changes: APIs are crucial in a hybrid cloud environment. As a consumer
of a provider's cloud services, an organization will rarely have any control over provider
changes to APIs. Cloud services that might have previously had compatible APIs may no
longer work. This is particularly a problem with AWS and OpenStack AWS-compatible
APIs. OpenStack was originally planned to maintain compatibility with changes in AWS
APIs. However, over time, the APIs have become more divergent in functionality. One
way to address this issue is to focus on using only the most common and basic APIs to
minimize potential conflicts.

Technical Considerations

A hybrid cloud environment requires inspection and understanding of technical issues that
are not only outside of an organization's data center, but potentially outside of an
organization's control. In many cases, it is necessary to ensure that the architecture and
CMP chosen can adapt to, not only to different environments, but also to the possibility of
change. In this situation, applications are crossing diverse platforms and are likely to be
located in diverse locations. All of these factors will influence and add complexity to the
design of a hybrid cloud architecture.

Incompatibilities

The only situation where cloud platform incompatibilities are not going to be an issue is
when working with clouds that are based on the same version and the same distribution of
OpenStack. Otherwise incompatibilities are virtually inevitable.
Incompatibility should be less of an issue for clouds that exclusively use the same version of
OpenStack, even if they use different distributions. The newer the distribution in question,
the less likely it is that there will be incompatibilities between version. This is due to the fact
that the OpenStack community has established an initiative to define core functions that
need to remain backward compatible between supported versions. The DefCore initiative
defines basic functions that every distribution must support in order to bear the name
"OpenStack".
Some vendors, however, add proprietary customizations to their distributions. If an
application or architecture makes use of these features, it will be difficult to migrate to or
use other types of environments. Anyone considering incorporating older versions of
OpenStack prior to Havana should consider carefully before attempting to incorporate
functionality between versions. Internal differences in older versions may be so great that
the best approach might be to consider the versions to be essentially diverse platforms, as
different as OpenStack and Amazon Web Services or Microsoft Azure.
The situation is more predictable if using different cloud platforms is incorporated from
inception. If the other clouds are not based on OpenStack, then all pretense of compatibility
vanishes, and CMP tools must account for the myriad of differences in the way operations
are handled and services are implemented. Some situations in which these incompatibilities
can arise include differences between the way in which a cloud:
Deploys instances
Manages networks

Treats applications
Implements services

Capacity planning

One of the primary reasons many organizations turn to a hybrid cloud system is to increase
capacity without having to make large capital investments. However, capacity planning is
still necessary when designing an OpenStack installation even if it is augmented with
external clouds.
Specifically, overall capacity and placement of workloads need to be accounted for when
designing for a mostly internally-operated cloud with the occasional capacity burs. The
long-term capacity plan for such a design needs to incorporate growth over time to prevent
the need to permanently burst into, and occupy, a potentially more expensive external
cloud. In order to avoid this scenario, account for the future applications and capacity
requirements and plan growth appropriately.
One of the drawbacks of capacity planning is unpredictability. It is difficult to predict the
amount of load a particular application might incur if the number of users fluctuates or the
application experiences an unexpected increase in popularity. It is possible to define
application requirements in terms of vCPU, RAM, bandwidth or other resources and plan
appropriately, but other clouds may not use the same metric or even the same
oversubscription rates.
Oversubscription is a method to emulate more capacity than they may physically be
present. For example, a physical hypervisor node with 32 gigabytes of RAM may host 24
instances, each provisioned with 2 gigabytes of RAM. As long as all 24 of them are not
concurrently utilizing 2 full gigabytes, this arrangement is a non-issue. However, some hosts
take oversubscription to extremes and, as a result, performance can frequently be
inconsistent. If at all possible, determine what the oversubscription rates of each host are
and plan capacity accordingly.

Security

The nature of a hybrid cloud environment removes complete control over the
infrastructure. Security becomes a stronger requirement because data or applications may
exist in a cloud that is outside of an organization's control. Security domains become an
important distinction when planning for a hybrid cloud environment and its capabilities. A
security domain comprises users, applications, servers or networks that share common

trust requirements and expectations within a system.
The security domains are:
1. Public
2. Guest
3. Management
4. Data
These security domains can be mapped individually to the organization's installation or
combined. For example, some deployment topologies combine both guest and data
domains onto one physical network, whereas other topologies may physically separate
these networks. In each case, the cloud operator should be aware of the appropriate
security concerns. Security domains should be mapped out against the specific OpenStack
deployment topology. The domains and their trust requirements depend upon whether the
cloud instance is public, private, or hybrid.
The public security domain is an entirely untrusted area of the cloud infrastructure. It can
refer to the Internet as a whole or simply to networks over which an organization has no
authority. This domain should always be considered untrusted. When considering hybrid
cloud deployments, any traffic traversing beyond and between the multiple clouds should
always be considered to reside in this security domain and is therefore untrusted.
Typically used for instance-to-instance traffic within a single datacenter, the guest security
domain handles compute data generated by instances on the cloud but not services that
support the operation of the cloud such as API calls. Public cloud providers that are used in a
hybrid cloud configuration which an organization does not control and private cloud
providers who do not have stringent controls on instance use or who allow unrestricted
internet access to instances should consider this domain to be untrusted. Private cloud
providers may consider this network as internal and therefore trusted only if there are
controls in place to assert that instances and tenants are trusted.
The management security domain is where services interact. Sometimes referred to as the
"control plane", the networks in this domain transport confidential data such as
configuration parameters, usernames, and passwords. In deployments behind an
organization's firewall, this domain is considered trusted. In a public cloud model which

could be part of an architecture, this would have to be assessed with the Public Cloud
provider to understand the controls in place.
The data security domain is concerned primarily with information pertaining to the storage
services within OpenStack. Much of the data that crosses this network has high integrity and
confidentiality requirements and depending on the type of deployment there may also be
strong availability requirements. The trust level of this network is heavily dependent on
deployment decisions and as such this is not assigned a default level of trust.
Consideration must be taken when managing the users of the system, whether operating or
utilizing public or private clouds. The identity service allows for LDAP to be part of the
authentication process. Including such systems in your OpenStack deployments may ease
user management if integrating into existing systems. Be mindful when utilizing 3rd party
clouds to explore authentication options applicable to the installation to help manage and
keep user authentication consistent.
Due to the process of passing usernames, passwords, and generated tokens between client
machines and API endpoints, placing API services behind hardware that performs SSL
termination is strongly recommended.
Within cloud components themselves, another component that needs security scrutiny is
the hypervisor. In a public cloud, organizations typically do not have control over the choice
of hypervisor. (Amazon uses its own particular version of Xen, for example.) In some cases,
hypervisors may be vulnerable to a type of attack called "hypervisor breakout" if they are
not properly secured. Hypervisor breakout describes the event of a compromised or
malicious instance breaking out of the resource controls of the hypervisor and gaining
access to the bare metal operating system and hardware resources.
If the security of instances is not considered important, there may not be an issue. In most
cases, however, enterprises need to avoid this kind of vulnerability, and the only way to do
that is to avoid a situation in which the instances are running on a public cloud. That does
not mean that there is a need to own all of the infrastructure on which an OpenStack
installation operates; it suggests avoiding situations in which hardware may be shared with
others.
There are other services worth considering that provide a bare metal instance instead of a
cloud. In other cases, it is possible to replicate a second private cloud by integrating with a
Private Cloud as a Service deployment, in which an organization does not buy hardware, but

also does not share it with other tenants. It is also possible use a provider that hosts a baremetal "public" cloud instance for which the hardware is dedicated only to one customer, or a
provider that offers Private Cloud as a Service.
Finally, it is important to realize that each cloud implements services differently. What keeps
data secure in one cloud may not do the same in another. Be sure to know the security
requirements of every cloud that handles the organization's data or workloads.
More information on OpenStack Security can be found at
http://docs.openstack.org/security-guide/

Utilization

When it comes to utilization, it is important that the CMP understands what workloads are
running, where they are running, and their preferred utilizations. For example, in most cases
it is desirable to run as many workloads internally as possible, utilizing other resources only
when necessary. On the other hand, situations exist in which the opposite is true. The
internal cloud may only be for development and stressing it is undesirable. In most cases, a
cost model of various scenarios helps with this decision, however this analysis is heavily
influenced by internal priorities. The important thing is the ability to efficiently make those
decisions on a programmatic basis.
The OpenStack Telemetry (Ceilometer) project is designed to provide information on the
usage of various OpenStack components. There are two limitations to consider: first, if there
is to be a large amount of data (for example, if monitoring a large cloud, or a very active one)
it is desirable to use a NoSQL backend for Ceilometer, such as MongoDB. Second, when
connecting to a non-OpenStack cloud, there will need to be a way to monitor that usage and
to provide that monitoring data back to the CMP.

Performance

Performance is of primary importance in the design of a cloud. When it comes to a hybrid
cloud deployment, many of the same issues for multi-site deployments apply, such as
network latency between sites. It is also important to think about the speed at which a
workload can be spun up in another cloud, and what can be done to reduce the time
necessary to accomplish that task. That may mean moving data closer to applications, or
conversely, applications closer to the data they process. It may mean grouping functionality
so that connections that require low latency take place over a single cloud rather than
spanning clouds. That may also mean ensuring that the CMP has the intelligence to know

which cloud can most efficiently run which types of workloads.
As with utilization, native OpenStack tools are available to assist. Ceilometer can measure
performance and, if necessary, OpenStack Orchestration via the Heat project can be used to
react to changes in demand by spinning up more resources. It is important to note, however,
that Orchestration requires special configurations in the client to enable functioning with
solution offerings from Amazon Web Services. When dealing with other types of clouds, it is
necessary to rely on the features of the CMP.

Components

The number and types of native OpenStack components that are available for use is
dependent on whether the deployment is exclusively an OpenStack cloud or not. If so, all of
the OpenStack components will be available for use, and in many ways the issues that need
to be considered will be similar to those that need to be considered for a multi-site
deployment.
That said, in any situation in which more than one cloud is being used, at least four
OpenStack tools will be considered:
OpenStack Compute (Nova): Regardless of deployment location, hypervisor choice
has a direct effect on how difficult it is to integrate with one or more additional clouds.
For example, integrating a Hyper-V based OpenStack cloud with Azure will have less
compatibility issues than if KVM is used.
Networking: Whether OpenStack Networking (Neutron) or Nova-network is used, the
network is one place where integration capabilities need to be understood in order to
connect between clouds.
OpenStack Telemetry (Ceilometer): Use of Ceilometer depends, in large part, on what
the other parts of the cloud are using.
OpenStack Orchestration (Heat): Similarly, Heat can be a valuable tool in orchestrating
tasks a CMP decides are necessary in an OpenStack-based cloud.

Special considerations

Hybrid cloud deployments also involve two more issues that are not common in other
situations:
Image portability: Note that, as of the Icehouse release, there is no single common image
format that is usable by all clouds. This means that images will need to be converted or
recreated when porting between clouds. To make things simpler, launch the smallest and

simplest images feasible, installing only what is necessary preferably using a deployment
manager such as Chef or Puppet. That means not to use golden images for speeding up the
process, however if the same images are being repeatedly deployed it may make more
sense to utilize this technique instead of provisioning applications on lighter images each
time.
API differences: The most profound issue that cannot be avoided when using a hybrid cloud
deployment with more than just OpenStack (or with different versions of OpenStack) is that
the APIs needed to perform certain functions are different. The CMP needs to know how to
handle all necessary versions. To get around this issue, some implementers build portals to
achieve a hybrid cloud environment, but a heavily developer-focused organization will get
more use out of a hybrid cloud broker SDK such as jClouds.

Operational Considerations

Hybrid cloud deployments present complex operational challenges. There are several
factors to consider that affect the way each cloud is deployed and how users and operators
will interact with each cloud. Not every cloud provider implements infrastructure
components the same way which may lead to incompatible interactions with workloads or a
specific Cloud Management Platform (CMP). Different cloud providers may also offer
different levels of integration with competing cloud offerings.

Monitoring

When selecting a CMP, one of the most important aspects to consider is monitoring. Gaining
valuable insight into each cloud is critical to gaining a holistic view of all involved clouds. In
choosing an existing CMP, determining whether it supports monitoring of all the clouds
involved or if compatible APIs are available which can be queried for the necessary
information, is vital. Once all the information about each cloud can be gathered and stored in
a searchable database, proper actions can be taken on that data offline so workloads will
not be impacted.

Agility

Implementing a hybrid cloud solution can provide application availability across disparate
cloud environments and technologies. This availability enables the deployment to survive a
complete disaster in any single cloud environment. Each cloud should provide the means to
quickly spin up new instances in the case of capacity issues or complete unavailability of a
single cloud installation.

Application Readiness

It is important to understand the type of application workloads that will be deployed across
the hybrid cloud environment. Enterprise workloads that depend on the underlying
infrastructure for availability are not designed to run on OpenStack. Although these types of
applications can run on an OpenStack cloud, if the application is not able to tolerate
infrastructure failures, it is likely to require significant operator intervention to recover.
Cloud workloads, however, are designed with fault tolerance in mind and the SLA of the
application is not tied to the underlying infrastructure. Ideally, cloud applications will be
designed to recover when entire racks and even data centers full of infrastructure
experience an outage.

Upgrades

OpenStack is a complex and constantly evolving collection of software. Upgrades may be
performed to one or more of the cloud environments involved. If a public cloud is involved in
the deployment, predicting upgrades may not be possible. Be sure to examine the
advertised SLA for any public cloud provider being used. Note that at massive scale, even
when dealing with a cloud that offers an SLA with a high percentage of uptime,
workloads must be able to recover at short notice.
Similarly, when upgrading private cloud deployments, care must be taken to minimize
disruption by making incremental changes and providing a facility to either rollback or
continue to roll forward when using a continuous delivery model.
Another consideration is upgrades to the CMP which may need to be completed in
coordination with any of the hybrid cloud upgrades. This may be necessary whenever API
changes are made in one of the cloud solutions in use to support the new functionality.

Network Operation Center (NOC)

When planning the Network Operation Center for a hybrid cloud environment, it is important
to recognize where control over each piece of infrastructure resides. If a significant portion
of the cloud is on externally managed systems, be prepared for situations in which it may
not be possible to make changes at all or at the most convenient time. Additionally,
situations of conflict may arise in which multiple providers have differing points of view on
the way infrastructure must be managed and exposed. This can lead to delays in root cause
and analysis where each insists the blame lies with the other provider.
It is important to ensure that the structure put in place enables connection of the
networking of both clouds to form an integrated system, keeping in mind the state of
handoffs. These handoffs must both be as reliable as possible and include as little latency as
possible to ensure the best performance of the overall system.

Maintainability

Operating hybrid clouds is a situation in which there is a greater reliance on third party
systems and processes. As a result of a lack of control of various pieces of a hybrid cloud
environment, it is not necessarily possible to guarantee proper maintenance of the overall
system. Instead, the user must be prepared to abandon workloads and spin them up again
in an improved state. Having a hybrid cloud deployment does, however, provide agility for
these situations by allowing the migration of workloads to alternative clouds in response to
cloud-specific issues.

Architecture
Logical Diagram

Once business and application requirements have been defined, the first step for designing a
hybrid cloud solution is to map out the dependencies between the expected workloads and
the diverse cloud infrastructures that need to support them. By mapping the applications
and the targeted cloud environments, a solution can be architected that enables the
broadest compatibility between cloud platforms and minimizes the need to create
workarounds and processes to fill identified gaps. Note the evaluation of the monitoring and
orchestration APIs available on each cloud platform and the relative levels of support for
them in the chosen Cloud Management Platform.

Image portability

The majority of cloud workloads currently run on instances using hypervisor technologies
such as KVM, Xen, or ESXi. The challenge is that each of these hypervisors use an image
format that is mostly, or not at all, compatible with one another. In a private or hybrid cloud
solution, this can be mitigated by standardizing on the same hypervisor and instance image
format but this is not always feasible. This is particularly evident if one of the clouds in the
architecture is a public cloud that is outside of the control of the designers.
There are conversion tools such as virt-v2v (http://libguestfs.org/virt-v2v/) and virt-edit
(http://libguestfs.org/virt-edit.1.html) that can be used in those scenarios but they are
often not suitable beyond very basic cloud instance specifications. An alternative is to build
a thin operating system image as the base for new instances. This facilitates rapid creation
of cloud instances using cloud orchestration or configuration management tools, driven by

the CMP, for more specific templating. Another more expensive option is to use a
commercial image migration tool. The issue of image portability is not just for a one time
migration. If the intention is to use the multiple cloud for disaster recovery, application
diversity or high availability, the images and instances are likely to be moved between the
different cloud platforms regularly.

Upper-Layer Services

Many clouds offer complementary services over and above the basic compute, network,
and storage components. These additional services are often used to simplify the
deployment and management of applications on a cloud platform.
Consideration is required to be given to moving workloads that may have upper-layer
service dependencies on the source cloud platform to a destination cloud platform that may
not have a comparable service. Conversely, the user can implement it in a different way or
by using a different technology. For example, moving an application that uses a NoSQL
database service such as MongoDB that is delivered as a service on the source cloud, to a
destination cloud that does not offer that service or may only use a relational database such
as MySQL, could cause difficulties in maintaining the application between the platforms.
There are a number of options that might be appropriate for the hybrid cloud use case:
Create a baseline of upper-layer services that are implemented across all of the cloud
platforms. For platforms that do not support a given service, create a service on top of
that platform and apply it to the workloads as they are launched on that cloud. For
example, OpenStack, via Trove, supports MySQL as a service but not NoSQL databases in
production. To move from or to run alongside on AWS a NoSQL workload would require
recreating the NoSQL database on top of OpenStack and automate the process of
implementing it using a tool such as OpenStack Orchestration (Heat).
Deploy a Platform as a Service (PaaS) technology such as Cloud Foundry or OpenShift
that abstracts the upper-layer services from the underlying cloud platform. The unit of
application deployment and migration is the PaaS and leverages the services of the PaaS
and only consumes the base infrastructure services of the cloud platform. The downside
to this approach is that the PaaS itself then potentially becomes a source of lock-in.
Use only the base infrastructure services that are common across all cloud platforms.
Use automation tools to create the required upper-layer services which are portable
across all cloud platforms. For example, instead of using any database services that are
inherent in the cloud platforms, launch cloud instances and deploy the databases on to
those instances using scripts or various configuration and application deployment tools.

Network Services

Network services functionality is a significant barrier for multiple cloud architectures. It
could be an important factor to assess when choosing a CMP and cloud provider.
Considerations are: functionality, security, scalability and High availability (HA). Verification
and ongoing testing of the critical features of the cloud endpoint used by the architecture
are important tasks.
Once the network functionality framework has been decided, a minimum functionality
test should be designed to confirm that the functionality is in fact compatible. This will
ensure testing and functionality persists during and after upgrades. Note that over time,
the diverse cloud platforms are likely to de-synchronize if care is not taken to maintain
compatibility. This is a particular issue with APIs.
Scalability across multiple cloud providers may dictate which underlying network
framework is chosen for the different cloud providers. It is important to have the
network API functions presented and to verify that the desired functionality persists
across all chosen cloud endpoint.
High availability (HA) implementations vary in functionality and design. Examples of
some common methods are Active-Hot-Standby, Active-Passive and Active-Active.
High availability and a test framework need to be developed to insure that the
functionality and limitations are well understood.
Security considerations, such as how data is secured between client and endpoint and
any traffic that traverses the multiple clouds, from eavesdropping to DoS activities must
be addressed. Business and regulatory requirements dictate the security approach that
needs to be taken.

Data

Replication has been the traditional method for protecting object store implementations. A
variety of different implementations have existed in storage architectures. Examples of this
are both synchronous and asynchronous mirroring. Most object stores and backend storage
systems have a method for replication that can be implemented at the storage subsystem
layer. Object stores also have implemented replication techniques that can be tailored to fit
a clouds needs. An organization must find the right balance between data integrity and data
availability. Replication strategy may also influence the disaster recovery methods
implemented.
Replication across different racks, data centers and geographical regions has led to the
increased focus of determining and ensuring data locality. The ability to guarantee data is

accessed from the nearest or fastest storage can be necessary for applications to perform
well. Examples of this are Hadoop running in a cloud. The user either runs with a native
HDFS, when applicable, or on a separate parallel file system such as those provided by
Hitachi and IBM. Special consideration should be taken when running embedded object
store methods to not cause extra data replication, which can create unnecessary
performance issues. Another example of ensuring data locality is by using Ceph. Ceph has a
data container abstraction called a pool. Pools can be created with replicas or erasure code.
Replica based pools can also have a ruleset defined to have data written to a “local” set of
hardware which would be the primary access and modification point.

Prescriptive Examples

Multi-cloud environments are typically created to facilitate these use cases:
Bursting workloads from private to public OpenStack clouds
Bursting workloads from private to public non-OpenStack clouds
High Availability across clouds (for technical diversity)
Examples of environments that address each of these use cases will be discussed in this
chapter.

Bursting to Public OpenStack Cloud

Company A's datacenter is running dangerously low on capacity. The option of expanding
the datacenter will not be possible in the foreseeable future. In order to accommodate the
continuously growing need for development resources in the organization, the decision was
make use of resource in the public cloud.
The company has an internal cloud management platform that will direct requests to the
appropriate cloud, depending on the currently local capacity.
This is a custom in-house application that has been written for this specific purpose.
An example for such a solution is described in the figure below.

This example shows two clouds, with a Cloud Management Platform (CMP) connecting

them. This guide does not attempt to cover a specific CMP, but describes how workloads are
typically orchestrated using the Orchestration and Telemetry services as shown in the
diagram above. It is also possibly to connect directly to the other OpenStack APIs with a
CMP.
The private cloud is an OpenStack cloud with one or more controllers and one or more
compute nodes. It includes metering provided by OpenStack Telemetry. As load increases
Telemetry captures this and the information is in turn processed by the CMP. As long as
capacity is available, the CMP uses the OpenStack API to call the Orchestration service to
create instances on the private cloud in response to user requests. When capacity is not
available on the private cloud, the CMP issues a request to the Orchestration service API of
the public cloud to create the instance on the public cloud.
In this example, the whole deployment was not directed to an external public cloud because
of the company's fear of lack of resource control and security concerns over control and
increased operational expense.
In addition, CompanyA has already established a datacenter with a substantial amount of
hardware, and migrating all the workloads out to a public cloud was not feasible.

Bursting to a Public non-OpenStack Cloud

Another common scenario is bursting workloads from the private cloud into a nonOpenStack public cloud such as Amazon Web Services (AWS) to take advantage of
additional capacity and scale applications as needed.
For an OpenStack-to-AWS hybrid cloud, the architecture looks similar to the figure below:

In this scenario CompanyA has an additional requirement in that the developers were
already using AWS for some of their work and did not want to change the cloud provider.
Primarily due to excessive overhead with network firewall rules that needed to be created
and corporate financial procedures that required entering into an agreement with a new
provider.
As long as the CMP is capable of connecting an external cloud provider with the appropriate
API, the workflow process will remain the same as the previous scenario. The actions the
CMP takes such as monitoring load, creating new instances, and so forth are the same, but
they would be performed in the public cloud using the appropriate API calls. For example, if
the public cloud is Amazon Web Services, the CMP would use the EC2 API to create a new
instance and assign an Elastic IP. That IP can then be added to HAProxy in the private cloud,
just as it was before. The CMP can also reference AWS-specific tools such as CloudWatch
and CloudFormation.
Several open source toolkits for building CMPs are now available that can handle this kind of
translation, including ManageIQ, jClouds, and JumpGate.

High Availability/Disaster Recovery

CompanyA has a requirement to be able to recover from catastrophic failure in their local
datacenter. Some of the workloads currently in use are running on their private OpenStack
cloud. Protecting the data involves block storage, object storage, and a database. The
architecture is designed to support the failure of large components of the system,

yet ensuring that the system will continue to deliver services. While the services remain
available to users, the failed components are restored in the background based on standard
best practice DR policies. To achieve the objectives, data is replicated to a second cloud, in a
geographically distant location. The logical diagram of the system is described in the figure
below:

This example includes two private OpenStack clouds connected with a Cloud Management
Platform (CMP). The source cloud, OpenStack Cloud 1, includes a controller and at least one
instance running MySQL. It also includes at least one block storage volume and one object
storage volume so that the data is available to the users at all times. The details of the
method for protecting each of these sources of data differs.
The object storage relies on the replication capabilities of the object storage provider.
OpenStack Object Storage is enabled so that it creates geographically
separated replicas that take advantage of this feature. It is configured so that at least one
replica exists in each cloud. In order to make this work a single array spanning both clouds is
configured with OpenStack Identity that uses Federated Identity and talks to both clouds,
communicating with OpenStack Object Storage through the Swift proxy.

For block storage, the replication is a little more difficult, and involves tools outside of
OpenStack itself. The OpenStack Block Storage volume is not set as the drive itself but as a
logical object that points to a physical backend. The disaster recovery is configured for Block
Storage for synchronous backup for the highest level of data protection, but asynchronous
backup could have been set as an alternative that is not as latency sensitive. For
asynchronous backup, the Cinder API makes it possible to export the data and also the
metadata of a particular volume, so that it can be moved and replicated elsewhere. More
information can be found here: https://blueprints.launchpad.net/cinder/+spec/cinderbackup-volume-metadata-support.
The synchronous backups create an identical volume in both clouds and chooses the
appropriate flavor so that each cloud has an identical backend. This was done by creating
volumes through the CMP, because the CMP knows to create identical volumes in both
clouds. Once this is configured, a solution, involving DRDB, is used to synchronize the actual
physical drives.
The database component is backed up using synchronous backups. MySQL does
not support geographically diverse replication, so disaster recovery is provided by replicating
the file itself. As it is not possible to use object storage as the back end of a database like
MySQL, Swift replication was not an option. It was decided not to store the data on another
geo-tiered storage system, such as Ceph, as block storage. This would have given another
layer of portection. Another option would have been to store the database on an OpenStack
Block Storage volume and backing it up just as any other block storage.

MASSIVELY SCALABLE

Introduction

A massively scalable architecture is defined as a cloud implementation that is either a very
large deployment, such as one that would be built by a commercial service provider, or one
that has the capability to support user requests for large amounts of cloud resources. An
example would be an infrastructure in which requests to service 500 instances or more at a
time is not uncommon. In a massively scalable infrastructure, such a request is fulfilled
without completely consuming all of the available cloud infrastructure resources. While the
high capital cost of implementing such a cloud architecture makes it cost prohibitive and is
only spearheaded by few organizations, many organizations are planning for massive
scalability moving toward the future.
A massively scalable OpenStack cloud design presents a unique set of challenges and
considerations. For the most part it is similar to a general purpose cloud architecture, as it is
built to address a non-specific range of potential use cases or functions. Typically, it is rare
that massively scalable clouds are designed or specialized for particular workloads. Like the
general purpose cloud, the massively scalable cloud is most often built as a platform for a
variety of workloads. Massively scalable OpenStack clouds are generally built as commercial
public cloud offerings since single private organizations rarely have the resources or need
for this scale.
Services provided by a massively scalable OpenStack cloud will include:
Virtual-machine disk image library
Raw block storage
File or object storage
Firewall functionality
Load balancing functionality
Private (non-routable) and public (floating) IP addresses
Virtualized network topologies
Software bundles
Virtual compte resources
Like a general purpose cloud, the instances deployed in a massively scalable OpenStack
cloud will not necessarily use any specific aspect of the cloud offering (compute, network, or
storage). As the cloud grows in scale, the scale of the number of workloads can cause stress
on all of the cloud components. Additional stresses are introduced to supporting
infrastructure including databases and message brokers. The architecture design for such a
cloud must account for these performance pressures without negatively impacting user

experience.

User Requirements

More so than other scenarios, defining user requirements for a massively scalable
OpenStack design architecture dictates approaching the design from two different, yet
sometimes opposing, perspectives: the cloud user, and the cloud operator. The expectations
and perceptions of the consumption and management of resources of a massively scalable
OpenStack cloud from the user point of view is distinctly different from that of the cloud
operator.

Legal Requirements

Many jurisdictions have legislative and regulatory requirements governing the storage and
management of data in cloud environments. Common areas of regulation include:
Data retention policies ensuring storage of persistent data and records management
to meet data archival requirements.
Data ownership policies governing the possession and responsibility for data.
Data sovereignty policies governing the storage of data in foreign countries or
otherwise separate jurisdictions.
Data compliance policies governing certain types of information needs to reside in
certain locations due to regular issues and, more importantly, cannot reside in other
locations for the same reason.
Examples of such legal frameworks include the data protection framework of the European
Union (http://ec.europa.eu/justice/data-protection/ ) and the requirements of the Financial
Industry Regulatory Authority (http://www.finra.org/Industry/Regulation/FINRARules/ ) in
the United States. Consult a local regulatory body for more information.

User Requirements

Massively scalable OpenStack clouds have the following user requirements:
The cloud user expects repeatable, dependable, and deterministic processes for
launching and deploying cloud resources. This could be delivered through a web-based
interface or publicly available API endpoints. All appropriate options for requesting cloud
resources need to be available through some type of user interface, a command-line
interface (CLI), or API endpoints.
Cloud users expect a fully self-service and on-demand consumption model. When an
OpenStack cloud reaches the "massively scalable" size, it means it is expected to be
consumed "as a service" in each and every way.
For a user of a massively scalable OpenStack public cloud, there will be no

expectations for control over security, performance, or availability. Only SLAs related to
uptime of API services are expected, and very basic SLAs expected of services offered.
The user understands it is his or her responsibility to address these issues on their own.
The exception to this expectation is the rare case of a massively scalable cloud
infrastructure built for a private or government organization that has specific
requirements.
As might be expected, the cloud user requirements or expectations that determine the
design are all focused on the consumption model. The user expects to be able to easily
consume cloud resources in an automated and deterministic way, without any need for
knowledge of the capacity, scalability, or other attributes of the cloud's underlying
infrastructure.

Operator Requirements

Whereas the cloud user should be completely unaware of the underlying infrastructure of
the cloud and its attributes, the operator must be able to build and support the
infrastructure, as well as how it needs to operate at scale. This presents a very demanding
set of requirements for building such a cloud from the operator's perspective:
First and foremost, everything must be capable of automation. From the deployment
of new hardware, compute hardware, storage hardware, or networking hardware, to the
installation and configuration of the supporting software, everything must be capable of
being automated. Manual processes will not suffice in a massively scalable OpenStack
design architecture.
The cloud operator requires that capital expenditure (CapEx) is minimized at all layers
of the stack. Operators of massively scalable OpenStack clouds require the use of
dependable commodity hardware and freely available open source software
components to reduce deployment costs and operational expenses. Initiatives like
OpenCompute (more information available at http://www.opencompute.org) provide
additional information and pointers. To cut costs, many operators sacrifice redundancy.
For example, redundant power supplies, redundant network connections, and redundant
rack switches.
Companies operating a massively scalable OpenStack cloud also require that
operational expenditures (OpEx) be minimized as much as possible. It is recommended
that cloud-optimized hardware is a good approach when managing operational
overhead. Some of the factors that need to be considered include power, cooling, and
the physical design of the chassis. It is possible to customize the hardware and systems
so they are optimized for this type of workload because of the scale of these

implementations.
Massively scalable OpenStack clouds require extensive metering and monitoring
functionality to maximize the operational efficiency by keeping the operator informed
about the status and state of the infrastructure. This includes full scale metering of the
hardware and software status. A corresponding framework of logging and alerting is also
required to store and allow operations to act upon the metrics provided by the metering
and monitoring solution(s). The cloud operator also needs a solution that uses the data
provided by the metering and monitoring solution to provide capacity planning and
capacity trending analysis.
A massively scalable OpenStack cloud will be a multi-site cloud. Therefore, the useroperator requirements for a multi-site OpenStack architecture design are also applicable
here. This includes various legal requirements for data storage, data placement, and data
retention; other jurisdictional legal or compliance requirements; image consistencyavailability; storage replication and availability (both block and file/object storage); and
authentication, authorization, and auditing (AAA), just to name a few. Refer to the "MultiSite" section for more details on requirements and considerations for multi-site
OpenStack clouds.
Considerations around physical facilities such as space, floor weight, rack height and
type, environmental considerations, power usage and power usage efficiency (PUE), and
physical security must also be addressed by the design architecture of a massively
scalable OpenStack cloud.

Technical Considerations
Foundations

Converting an existing OpenStack environment that was designed for a different purpose to
be massively scalable is a formidable task. When building a massively scalable environment
from the ground up, make sure the initial deployment is built with the same principles and
choices that apply as the environment grows. For example, a good approach is to deploy the
first site as a multi-site environment. This allows the same deployment and segregation
methods to be used as the environment grows to separate locations across dedicated links
or wide area networks. In a hyperscale cloud, scale trumps redundancy. Applications must
be modified with this in mind, relying on the scale and homogeneity of the environment to
provide reliability rather than redundant infrastructure provided by non-commodity
hardware solutions.

Infrastructure Segregation

Fortunately, OpenStack services are designed to support massive horizontal scale. Be aware
that this is not the case for the entire supporting infrastructure. This is particularly a problem
for the database management systems and message queues used by the various
OpenStack services for data storage and remote procedure call communications.
Traditional clustering techniques are typically used to provide high availability and some
additional scale for these environments. In the quest for massive scale, however, additional
steps need to be taken to relieve the performance pressure on these components to
prevent them from negatively impacting the overall performance of the environment. It is
important to make sure that all the components are in balance so that, if and when the
massively scalable environment fails, all the components are at, or close to, maximum
capacity.
Regions are used to segregate completely independent installations linked only by an
Identity and Dashboard (optional) installation. Services are installed with separate API
endpoints for each region, complete with separate database and queue installations. This
exposes some awareness of the environment's fault domains to users and gives them the
ability to ensure some degree of application resiliency while also imposing the requirement
to specify which region their actions must be applied to.
Environments operating at massive scale typically need their regions or sites subdivided
further without exposing the requirement to specify the failure domain to the user. This

provides the ability to further divide the installation into failure domains while also providing
a logical unit for maintenance and the addition of new hardware. At hyperscale, instead of
adding single compute nodes, administrators may add entire racks or even groups of racks
at a time with each new addition of nodes exposed via one of the segregation concepts
mentioned herein.

Cells

Cells provide the ability to subdivide the compute portion of an OpenStack installation,
including regions, while still exposing a single endpoint. In each region an API cell is created
along with a number of compute cells where the workloads actually run. Each cell gets its
own database and message queue setup (ideally clustered), providing the ability to
subdivide the load on these subsystems, improving overall performance.
Within each compute cell a complete compute installation is provided, complete with full
database and queue installations, scheduler, conductor, and multiple compute hosts. The
cells scheduler handles placement of user requests from the single API endpoint to a
specific cell from those available. The normal filter scheduler then handles placement within
the cell.
The downside of using cells is that they are not well supported by any of the OpenStack
services other than compute. Also, they do not adequately support some relatively standard
OpenStack functionality such as security groups and host aggregates. Due to their relative
newness and specialized use, they receive relatively little testing in the OpenStack gate.
Despite these issues, however, cells are used in some very well known OpenStack
installations operating at massive scale including those at CERN and Rackspace.

Host Aggregates

Host Aggregates enable partitioning of OpenStack Compute deployments into logical
groups for load balancing and instance distribution. Host aggregates may also be used to
further partition an availability zone. Consider a cloud which might use host aggregates to
partition an availability zone into groups of hosts that either share common resources, such
as storage and network, or have a special property, such as trusted computing hardware.
Host aggregates are not explicitly user targetable; instead they are implicitly targeted via the
selection of instance flavors with extra specifications that map to host aggregate metadata.

Availability Zones

Availability zones provide another mechanism for subdividing an installation or region. They
are, in effect, Host aggregates that are exposed for (optional) explicit targeting by users.

Unlike cells, they do not have their own database server or queue broker but simply
represent an arbitrary grouping of compute nodes. Typically, grouping of nodes into
availability zones is based on a shared failure domain based on a physical characteristic such
as a shared power source, physical network connection, and so on. Availability Zones are
exposed to the user because they can be targeted; however, users are not required to target
them. An alternate approach is for the operator to set a default availability zone to schedule
instances to other than the default availability zone of nova.

Segregation Example

In this example the cloud is divided into two regions, one for each site, with two availability
zones in each based on the power layout of the data centers. A number of host aggregates
have also been defined to allow targeting of virtual machine instances using flavors, that
require special capabilities shared by the target hosts such as SSDs, 10 G networks, or GPU
cards.

Operational Considerations
Automation

In order to run at massive scale, it is important to plan on the automation of as many of the
operational processes as possible. Automation includes the configuration of provisioning,
monitoring and alerting systems. Part of the automation process includes the capability to
determine when human intervention is required and who should act. The objective is to
increase the ratio of operational staff to running systems as much as possible to reduce
maintenance costs. In a massively scaled environment, it is impossible for staff to give each
system individual care.
Configuration management tools such as Puppet or Chef allow operations staff to
categorize systems into groups based on their role and thus create configurations and
system states that are enforced through the provisioning system. Systems that fall out of
the defined state due to errors or failures are quickly removed from the pool of active nodes
and replaced.
At large scale the resource cost of diagnosing individual systems that have failed is far
greater than the cost of replacement. It is more economical to immediately replace the
system with a new system that can be provisioned and configured automatically and quickly
brought back into the pool of active nodes. By automating tasks that are labor intensive,
repetitive, and critical to operations with automation, cloud operations teams are able to be
managed more efficiently because fewer resources are needed for these babysitting tasks.
Administrators are then free to tackle tasks that cannot be easily automated and have
longer-term impacts on the business such as capacity planning.

The Bleeding Edge

Running OpenStack at massive scale requires striking a balance between stability and
features. For example, it might be tempting to run an older stable release branch of
OpenStack to make deployments easier. However, when running at massive scale, known
issues that may be of some concern or only have minimal impact in smaller
deployments could become pain points at massive scale. If the issue is well known, in many
cases, it may be resolved in more recent releases. The OpenStack community can help
resolve any issues reported by the applying the collective expertise of the OpenStack
developers.
When issues crop up, the number of organizations running at a similar scale is a relatively

tiny proportion of the OpenStack community, therefore it is important to share these issues
with the community and be a vocal advocate for resolving them. Some issues only manifest
when operating at large scale and the number of organizations able to duplicate and
validate an issue is small, so it will be important to document and dedicate resources to their
resolution.
In some cases, the resolution to the problem is ultimately to deploy a more recent version of
OpenStack. Alternatively, when the issue needs to be resolved in a production environment
where rebuilding the entire environment is not an option, it is possible to deploy just the
more recent separate underlying components required to resolve issues or gain significant
performance improvements. At first glance, this could be perceived as potentially exposing
the deployment to increased risk to and instability. However, in many cases it could be an
issue that has not been discovered yet.
It is advisable to cultivate a development and operations organization that is responsible for
creating desired features, diagnose and resolve issues, and also build the infrastructure for
large scale continuous integration tests and continuous deployment. This helps catch bugs
early and make deployments quicker and less painful. In addition to development resources,
the recruitment of experts in the fields of message queues, databases, distributed systems,
and networking, cloud and storage is also advisable.

Growth and Capacity Planning

An important consideration in running at massive scale is projecting growth and utilization
trends to plan capital expenditures for the near and long term. Utilization metrics for
compute, network, and storage as well as a historical record of these metrics are required.
While securing major anchor tenants can lead to rapid jumps in the utilization rates of all
resources, the steady adoption of the cloud inside an organizations or by public consumers
in a public offering will also create a steady trend of increased utilization.

Skills and Training

Projecting growth for storage, networking, and compute is only one aspect of a growth plan
for running OpenStack at massive scale. Growing and nurturing development and
operational staff is an additional consideration. Sending team members to OpenStack
conferences, meetup events, and encouraging active participation in the mailing lists and
committees is a very important way to maintain skills and forge relationships in the
community. A list of OpenStack training providers in the marketplace can be found here:
http://www.openstack.org/marketplace/training/.

SPECIALIZED CASES

Introduction

Although most OpenStack architecture designs fall into one of the seven major scenarios
outlined in other sections (compute focused, network focused, storage focused, general
purpose, multi-site, hybrid cloud, and massively scalable), there are a few other use cases
that are unique enough they can't be neatly categorized into one of the other major sections.
This section discusses some of these unique use cases with some additional details and
design considerations for each use case.
Specialized Networking: This describes running networking-oriented software that may
involve reading packets directly from the wire or participating in routing protocols.
Software-Defined Networking: This use case details both running an SDN controller
from within OpenStack as well as participating in a software-defined network.
Desktop as a Service: This is for organizations that want to run a virtualized desktop
environment on a cloud. This can apply to private or public clouds.
OpenStack on OpenStack: Some organizations are finding that it makes technical sense to
build a multi-tiered cloud by running OpenStack on top of an OpenStack installation.
Specialized Hardware: Some highly specialized situations will require the use
of specialized hardware devices from within the OpenStack environment.

Multi-Hypervisor Example

A financial company requires a migration of its applications from a traditional virtualized
environment to an API driven, orchestrated environment. A number of their applications
have strict support requirements which limit what hypervisors they are supported on,
however the rest do not have such restrictions and do not need the same features. Because
of these requirements, the overall target environment needs multiple hypervisors.
The current environment consists of a vSphere environment with 20 VMware ESXi
hypervisors supporting 300 instances of various sizes. Approximately 50 of the instances
must be run on ESXi but the rest have more flexible requirements.
The company has decided to bring the management of the overall system into a common
platform provided by OpenStack.

The approach is to run a host aggregate consisting of KVM hypervisors for the general

purpose instances and a separate host aggregate for instances requiring ESXi. This way,
workloads that must be run on ESXi can be targeted at those hypervisors, but the rest can
be targeted at the KVM hypervisors.
Images in the OpenStack Image Service have particular hypervisor metadata attached so
that when a user requests a certain image, the instance will spawn on the relevant
aggregate. Images for ESXi are stored in VMDK format. QEMU disk images can be converted
to VMDK, VMFS Flat Disks, which includes thin, thick, zeroed-thick, and eager-zeroed-thick.
Note that once a VMFS thin disk is exported from VMFS to a non-VMFS location, like the
OpenStack Image Service, it becomes a preallocated flat disk. This impacts the transfer time
from the OpenStack Image Service to the data store when the full preallocated flat disk,
rather than the thin disk, must be transferred.
This example has the additional complication that, rather than being spawned directly on a
hypervisor simply by calling a specific host aggregate using the metadata of the image, the
VMware host aggregate compute nodes communicate with vCenter which then
requests that the instance be scheduled to run on an ESXi hypervisor. As of the Icehouse
release, this functionality requires that VMware Distributed Resource Scheduler (DRS) be
enabled on a cluster and set to "Fully Automated".
Due to the DRS requirement, note that vSphere requires shared storage (DRS uses vMotion,
which requires shared storage). The solution uses shared storage to provide Block Storage
capabilities to the KVM instances while also providing the storage for vSphere. The
environment uses a dedicated data network to provide this functionality, therefore the
compute hosts should have dedicated NICs to support this dedicated traffic. vSphere
supports the use of OpenStack Block Storage to present storage from a VMFS datastore to
an instance, so the use of Block Storage in this architecture supports both hypervisors.
In this case, network connectivity is provided by OpenStack Networking with the VMware
NSX plugin driver configured. Alternatively, the system could use nova-network, which is
supported by both hypervisors used in this design, but has limitations. Specifically, security
groups are not supported on vSphere with nova-network. With VMware NSX as part of the
design, however, when a user launches an instance within either of the host aggregates, the
instances are attached to appropriate network overlay-based logical networks as defined by
the user.
Note that care must be taken with this approach, as there are design considerations around
the OpenStack Compute integration. When using vSphere with OpenStack, the nova-

compute service that is configured to communicate with vCenter shows up as a single large
hypervisor representing the entire ESXi cluster (multiple instances of nova-compute can be
run to represent multiple ESXi clusters or to connect to multiple vCenter servers). If the
process running the nova-compute service crashes, the connection to that particular
vCenter Server-and any ESXi clusters behind it-are severed and it will not be possible to
provision more instances on that vCenter, despite the fact that vSphere itself could be
configured for high availability. Therefore, it is important to monitor the nova-compute
service that connects to vSphere for any disruptions.

Specialized Networking Example
Use Case

Some applications that interact with a network require more specialized connectivity.
Applications such as a looking glass require the ability to connect to a BGP peer, or route
participant applications may need to join a network at a layer 2 level.

Challenges

Connecting specialized network applications to their required resources alters the design of
an OpenStack installation. Installations that rely on overlay networks are unable to support
a routing participant, and may also block layer 2 listeners.

Possible Solutions

Deploying an OpenStack installation using Neutron with a provider network will allow direct
layer 2 connectivity to an upstream networking device. This design provides the layer 2
connectivity required to communicate via Intermediate System-to-Intermediate System
(ISIS) protocol or to pass packets controlled via an OpenFlow controller. Using the multiple
layer 2 plugin with an agent such as Open vSwitch would allow a private connection through
a VLAN directly to a specific port in a layer 3 device. This would allow a BGP point to point
link to exist that will join the autonomous system. Avoid using layer 3 plugins as they will
divide the broadcast domain and prevent router adjacencies from forming.

Software Defined Networking (SDN)
Use Case

Software Defined Networking is the separation of the data plane and control plane. SDN has
become a popular method of managing and controlling packet flows within networks. SDN
uses overlays or directly controlled layer 2 devices to determine flow paths, and as such
presents challenges to a cloud environment. Some designers may wish to run their
controllers within an OpenStack installation. Others may wish to have their installations
participate in an SDN-controlled network.

Challenges

SDN is a relatively new concept that is not yet standardized, so SDN systems come in a
variety of different implementations. Because of this, a truly prescriptive architecture is not
feasible. Instead, examine the differences between an existing or intended OpenStack design
and determine where the potential conflict and gaps can be found.

Possible Solutions

If an SDN implementation requires layer 2 access because it directly manipulates switches,
then running an overlay network or a layer 3 agent may not be advisable. If the controller
resides within an OpenStack installation, it may be necessary to build an ML2 plugin and
schedule the controller instances to connect to tenant VLANs that then talk directly to the
switch hardware. Alternatively, depending on the external device support, use a tunnel that
terminates at the switch hardware itself.

Diagram

OpenStack hosted SDN controller:

OpenStack participating in an SDN controller network:

Desktop as a Service
Use Case

Virtual Desktop Infrastructure (VDI) is a service that hosts user desktop environments on
remote servers. This application is very sensitive to network latency and requires a high
performance compute environment. Traditionally these types of environments have not
been put on cloud environments because few clouds are built to support such a demanding
workload that is so exposed to end users. Recently, as cloud environments become more
robust, vendors are starting to provide services that allow virtual desktops to be hosted in
the cloud. In the not too distant future, OpenStack could be used as the underlying
infrastructure to run a virtual infrastructure environment, either in-house or in the cloud.

Challenges
Planning for performance

Designing an infrastructure that is suitable to host virtual desktops is a very different task to
that of most virtual workloads. The infrastructure will need to be designed, for example:
Boot storms - What happens when hundreds or thousands of users log in during shift
changes, affects the storage design.
The performance of the applications running in these virtual desktops
Operating system and compatibility with the OpenStack hypervisor

Broker

The Connection Broker is a central component of the architecture that determines which
Remote Desktop Host will be assigned or connected to the user. The broker is often a fullblown management product allowing for the automated deployment and provisioning of
Remote Desktop Hosts.

Possible Solutions

There a number of commercial products available today that provide such a broker solution
but nothing that is native in the OpenStack project. There of course is also the option of not
providing a broker and managing this manually - but this would not suffice as a large scale,
enterprise solution.

Diagram

OpenStack on OpenStack
Use Case

In some cases it is necessary to run OpenStack nested on top of another OpenStack cloud.
This scenario allows for complete OpenStack cloud environments to be managed and
provisioned on instances running on hypervisors and servers controlled by the underlying
OpenStack cloud. Public cloud providers can use this technique to effectively manage the
upgrade and maintenance process on complete OpenStack-based clouds. Developers and
those testing OpenStack can also use the guidance to provision their own OpenStack
environments on available OpenStack Compute resources, whether Public or Private.

Challenges
Network

The network aspect of deploying a nested cloud is the most complicated aspect of this
architecture. When using VLANs, these will need to be exposed to the physical ports on
which the undercloud runs, as the bare metal cloud owns all the hardware, but they also
need to be exposed to the nested levels as well. Alternatively, network overlay technologies
can be used on the overcloud (the OpenStack cloud running on OpenStack) to provide the
required software defined networking for the deployment.

Hypervisor

A key question to address in this scenario is the decision about which approach should be
taken to provide a nested hypervisor in OpenStack. This decision influences which operating
systems can be used for the deployment of the nested OpenStack deployments.

Possible Solutions
Deployment

Deployment of a full stack can be challenging but this difficulty can be readily be mitigated
by creating a Heat template to deploy the entire stack or a configuration management
system. Once the Heat template is created, deploying additional stacks will be a trivial thing
and can be performed in an automated fashion.
The OpenStack-On-OpenStack project (TripleO) is addressing this issue - although at the
current time the project does not provide comprehensive coverage for the nested stacks.
More information can be found at https://wiki.openstack.org/wiki/TripleO.

Hypervisor

In the case of running TripleO, the underlying OpenStack cloud deploys the Compute nodes
as bare-metal. OpenStack would then be deployed on these Compute bare-metal servers
with the appropriate hypervisor, such as KVM.
In the case of running smaller OpenStack clouds for testing purposes, and performance
would not be a critical factor, QEMU can be utilized instead. It is also possible to run a KVM
hypervisor in an instance (http://davejingtian.org/2014/03/30/nested-kvm-just-for-fun/),
though this is not a supported configuration, and could be a complex solution for such a use
case.

Diagram

Specialized Hardware
Use Cases

Certain workloads require specialized hardware devices that are either difficult to virtualize
or impossible to share. Applications such as load balancers, highly parallel brute force
computing, and direct to wire networking may need capabilities that basic OpenStack
components do not provide.

Challenges

Some applications need access to hardware devices to either improve performance or
provide capabilities that are not virtual CPU, RAM, network or storage. These can be a shared
resource, such as a cryptography processor, or a dedicated resource such as a Graphics
Processing Unit. OpenStack has ways of providing some of these, while others may need
extra work.

Solutions

In order to provide cryptography offloading to a set of instances, it is possible to use Glance
configuration options to assign the cryptography chip to a device node in the guest. The
documentation at http://docs.openstack.org/cli-reference/content/chapter_cli-glanceproperty.html contains further information on configuring this solution, but it allows all
guests using the configured images to access the hypervisor cryptography device.

If direct access to a specific device is required, it can be dedicated to a single instance per
hypervisor through the use of PCI pass-through. The OpenStack administrator needs to
define a flavor that specifically has the PCI device in order to properly schedule instances.
More information regarding PCI pass-through, including instructions for implementing and
using it, is available
at https://wiki.openstack.org/wiki/Pci_passthrough#How_to_check_PCI_status_with_PCI_

References

Data Protection framework of the European Union: http://ec.europa.eu/justice/dataprotection/
Guidance on Data Protection laws governed by the EU
Depletion of IPv4
Addresses: http://www.internetsociety.org/deploy360/blog/2014/05/goodbye-ipv4-ianastarts-allocating-final-address-blocks/
Article describing how IPv4 addresses and the migration to IPv6 is inevatible
Ethernet Switch
Reliability: http://www.garrettcom.com/techsupport/papers/ethernet_switch_reliability.pd
f
Research whitepaper on Ethernet Switch reliability
Financial Industry Regulatory
Authority: http://www.finra.org/Industry/Regulation/FINRARules/
Requirements of the Financial Industry Regulatory Authority in the USA
Image Service property keys: http://docs.openstack.org/cli-reference/content/chapter_cliglance-property.html
Glance API property keys allows the administrator to attach custom characteristics
to images
LibGuestFS Documentation: http://libguestfs.org
Official LibGuestFS documentation
Logging and Monitoring http://docs.openstack.org/openstackops/content/logging_monitoring.html
Official OpenStack Operations documentation
ManageIQ Cloud Management Platform: http://manageiq.org/
An Open Source Cloud Management Platform for managing multiple clouds
N-Tron Network Availability: http://www.n-tron.com/pdf/network_availability.pdf
Research whitepaper on network availability
Nested KVM: http://davejingtian.org/2014/03/30/nested-kvm-just-for-fun
Blog Post on how to nest KVM under KVM.

Open Compute Project: http://www.opencompute.org/
The Open Compute Project Foundation’s mission is to design and enable the delivery of the
most efficient server, storage and data center hardware designs for scalable computing.
OpenStack Flavors: http://docs.openstack.org/openstack-ops/content/flavors.html
Official OpenStack documentation
OpenStack High Availability Guide: http://docs.openstack.org/high-availabilityguide/content/
Information on how to provide redundancy for the OpenStack components
OpenStack Hypervisor Support
Matrix: https://wiki.openstack.org/wiki/HypervisorSupportMatrix
Matrix of supported hypervisors and capabilities when used with OpenStack
OpenStack Object Store (Swift) Replication
Reference: http://docs.openstack.org/developer/swift/replication_network.html
Developer documentation of Swift replication
OpenStack Operations Guide: http://docs.openstack.org/openstack-ops/
The OpenStack Operations Guide provides information on setting up and installing
OpenStack
OpenStack Security Guide: http://docs.openstack.org/security-guide/
The OpenStack Security Guide provides information on securing OpenStack deployments
OpenStack Training Marketplace: http://www.openstack.org/marketplace/training
The OpenStack Market for training and Vendors providing training on OpenStack.
PCI passthrough:
https://wiki.openstack.org/wiki/Pci_passthrough#How_to_check_PCI_status_with_PCI_
api_paches
The PCI api patches extends the servers/os-hypervisor to show PCI information for instance
and compute node, and also provides a resource endpoint to show PCI information.
TripleO: https://wiki.openstack.org/wiki/TripleO
TripleO is a program aimed at installing, upgrading and operating OpenStack clouds using
OpenStack's own cloud facilities as the foundation.

Glossary

6to4 - A mechanism that allows IPv6 packets to be transmitted over an IPv4 network,
providing a strategy for migrating to IPv6.
AAA - authentication, authorization and auditing.
Anycast - A network routing methodology that routes traffic from a single sender to the
nearest node, in a pool of nodes.
ARP - Address Resolution Protocol - the protocol by which layer 3 IP addresses are
resolved into layer 2, link local addresses.
BGP - Border Gateway Protocol is a dynamic routing protocol that connects autonomous
systems together. Considered the backbone of the Internet, this protocol connects disparate
networks together to form a larger network.
Boot Storm - When hundreds of users log in to and consume resources at the same time,
causing significant performance degradation. This problem is particularly common in Virtual
Desktop Infrastructure (VDI) environments.
Broadcast Domain - The layer 2 segment shared by a group of network connected
nodes.
Bursting - The practice of utilizing a secondary environment to elastically build instances
on-demand when the primary environment is resource constrained.
Capital Expenditure (CapEx) - A capital expense, capital expenditure, CapEx is an initial
cost for building a product, business, or system.
Cascading Failure - A scenario where a single failure in a system creates a cascading
effect, where other systems fail as load is transferred from the failing system.
CDN - Content delivery network - a specialized network that is used to distribute content
to clients, typically located close to the client for increased performance.
Cells - An OpenStack Compute (Nova) feature, where a compute deployment can be split
into smaller clusters or cells with their own queue and database for performance and
scalability, while still providing a single API endpoint.
CI/CD - Continuous Integration / Continuous Deployment, a methodology where
software is continually built and unit tests run for each change that is merged, or proposed
for merge. Continuous Deployment is a software development methodology where changes
are deployed into production as they are merged into source control, rather than being
collected into a release and deployed at regular intervals
Cloud Broker - A cloud broker is a third-party individual or business that acts as an
intermediary between the purchaser of a cloud computing service and the sellers of that
service. In general, a broker is someone who acts as an intermediary between two or more
parties during negotiations.

Cloud Consumer - User that consumes cloud instances, storage, or other resources in a
cloud environment. This user interacts with OpenStack or other cloud management tools.
Cloud Management Platform (CMP) - Products that provide a common interface to
manage multiple cloud environments or platforms.
Connection Broker - In desktop virtualization, a connection broker is a software program
that allows the end-user to connect to an available desktop.
Direct Attached Storage (DAS) - Data storage that is directly connected to a machine.
DefCore - DefCore sets base requirements by defining capabilities, code and must-pass
tests for all OpenStack products. This definition uses community resources and involvement
to drive interoperability by creating the minimum standards for products labeled
"OpenStack." See https://wiki.openstack.org/wiki/Governance/CoreDefinition for more
information.
Desktop as a Service (DaaS) - A platform that provides a suite of desktop environments
that users may log in to to receive a desktop experience from any location. This may provide
general use, development, or even homogenous testing environments.
Direct Server Return - A technique in load balancing where an initial request is routed
through a load balancer, and the reply is sent from the responding node directly to the
requester.
Denial of Service (DoS) - In computing, a denial-of-service or distributed denial-ofservice attack is an attempt to make a machine or network resource unavailable to its
intended users.
Distributed Replicated Block Device (DRBD) - The Distributed Replicated Block Device
(DRBD) is a distributed replicated storage system for the Linux platform.
Differentiated Service Code Point (DSCP) - Defined in RFC 2474, this field in IPv4 and
IPv6 headers is used to define classes of network traffic, for quality of service purposes.
External Border Gateway Protocol (eBGP) - External Border Gateway Protocol describes a
specific implementation of BGP designed for inter-autonomous system communication
Elastic IP - An Amazon Web Services concept, which is an IP address that can be
dynamically allocated and re-assigned to running instances on the fly. The OpenStack
equivalent is a Floating IP.
Encapsulation - The practice of placing one packet type within another for the purposes
of abstracting or securing data. Examples include GRE, MPLS, or IPSEC.
External Cloud - A cloud environment that exists outside of the control of an
organization. Referred to for hybrid cloud to indicate a public cloud or an off-site hosted
cloud.
Federated Cloud - A federated cloud describes a multiple sets of cloud resources, for
example compute or storage, that are managed by a centralized endpoint.

Flow - A series of packets that are stateful in nature and represent a session. Usually
represented by a TCP stream, but can also indicate other packet types that when combined
comprise a connection between two points.
Golden Image - An operating system image that contains a set of pre-installed software
packages and configurations. This may be used to build standardized instances that have
the same base set of configuration to improve mean time to functional application
Graphics Processing Unit (GPU) - A single chip processor with integrated transform,
lighting, triangle setup/clipping, and rendering engines that is capable of processing a
minimum of 10 million polygons per second. Traditional uses are any compute problem
that can be represented as a vector or matrix operation.
Hadoop Distributed File System (HDFS) - A distributed file-system that stores data on
commodity machines, providing very high aggregate bandwidth across the cluster.
High Availability (HA) - High availability system design approach and associated service
implementation that ensures a prearranged level of operational performance will be met
during a contractual measurement period.
High Performance Computing (HPC) - Also known as distributed computing - used for
computation intensive processes run on a large number of instances
Hierarchical Storage Management (HSM) - Hierarchical storage management is a data
storage technique, which automatically moves data between high-cost and low-cost
storage media
Hot Standby Router Protocol (HSRP) - Hot Standby Router Protocol is a Cisco proprietary
redundancy protocol for establishing a fault-tolerant default gateway, and has been
described in detail in RFC 2281.
Hybrid Cloud - Hybrid cloud is a composition of two or more clouds (private, community
or public) that remain distinct entities but are bound together, offering the benefits of
multiple deployment models. Hybrid cloud can also mean the ability to connect colocation,
managed and/or dedicated services with cloud resources.
Interior Border Gateway Protocol (iBGP) - Interior Border Gateway Protocol is a an
interior gateway protocol designed to exchange routing and reachability information within
autonomous systems.
Interior Gateway Protocol (IGP) - An Interior Gateway Protocol is a type of protocol used
for exchanging routing information between gateways (commonly routers) within an
Autonomous System (for example, a system of corporate local area networks). This routing
information can then be used to route network-level protocols like IP.
Input/Output Operations Per Second (IOPS) - A common performance measurement
used to benchmark computer storage devices like hard disk drives, solid state drives, and
storage area networks.

jClouds - An open source multi-cloud toolkit for the Java platform that gives you the
freedom to create applications that are portable across clouds while giving you full control
to use cloud-specific features.
Jitter - Is the deviation from true periodicity of a presumed periodic signal in electronics
and telecommunications, often in relation to a reference clock source.
Jumbo Frame - Ethernet frames with more than 1500 bytes of payload.
Kernel-based Virtual Machine (KVM) - A full virtualization solution for Linux on x86
hardware containing virtualization extensions (Intel VT or AMD-V). It consists of a loadable
kernel module, that provides the core virtualization infrastructure and a processor specific
module.
LAG - Link aggregation group is a term to describe various methods of combining
(aggregating) multiple network connections in parallel into a group to increase throughput
beyond what a single connection could sustain, and to provide redundancy in case one of
the links fail.
Layer 2 - The data link layer provides a reliable link between two directly connected
nodes, by detecting and possibly correcting errors that may occur in the physical layer.
Layer 3 - The network layer provides the functional and procedural means of transferring
variable length data sequences (called datagrams) from one node to another connected to
the same network.
Legacy System - An old method, technology, computer system, or application program
that is considered outdated.
Looking Glass - A tool that provides information on backbone routing and network
efficiency.
Microsoft Azure - A cloud computing platform and infrastructure, created by Microsoft,
for building, deploying and managing applications and services through a global network of
Microsoft-managed datacenters.
MongoDB - A cross-platform document-oriented database. Classified as a NoSQL
database, MongoDB eschews the traditional table-based relational database structure in
favor of JSON-like documents with dynamic schemas.
Mean Time Before Failures (MTBF) - Mean time before failures is the predicted elapsed
time before inherent failures of a system during operation. MTBF can be calculated as the
arithmetic mean (average) time between failures of a system.
Maximum Transmission Unit (MTU) - The maximum transmission unit of a
communications protocol of a layer is the size (in bytes) of the largest protocol data unit that
the layer can pass onwards.
NAT64 - NAT64 is a mechanism to allow IPv6 hosts to communicate with IPv4 servers.
The NAT64 server is the endpoint for at least one IPv4 address and an IPv6 network

segment of 32-bits.
Network Functions Virtualization (NFV) - Network Functions Virtualization is a network
architecture concept that proposes using IT virtualization related technologies, to virtualize
entire classes of network node functions into building blocks that may be connected, or
chained, together to create communication services.
NoSQL - A NoSQL or Not Only SQL database provides a mechanism for storage and
retrieval of data that is modeled in means other than the tabular relations used in relational
databases.
Open vSwitch - Open vSwitch is a production quality, multilayer virtual switch licensed
under the open source Apache 2.0 license. It is designed to enable massive network
automation through programmatic extension, while still supporting standard management
interfaces and protocols (e.g. NetFlow, sFlow, SPAN, RSPAN, CLI, LACP, 802.1ag).
Operational Expenditure (OPEX) - An operating expense, operating expenditure,
operational expense, operational expenditure or OPEX is an ongoing cost for running a
product, business, or system.
Original Design Manufacturers (ODM) - Original Design Manufacturers, a company which
designs and manufactures a product which is specified and eventually branded by another
firm for sale.
Overlay Network - An overlay network is a computer network which is built on the top of
another network. Nodes in the overlay can be thought of as being connected by virtual or
logical links, each of which corresponds to a path, perhaps through many physical links, in
the underlying network.
Packet Storm - A cause of degregated service or failure that occurs when a network
system is overwhelmed by continuous multicast or broadcast traffic.
Platform as a Service (PaaS) - Platform as a Service is a category of cloud computing
services that provides a computing platform and a solution stack as a service.
Power Usage Effectiveness (PUE) - Power usage effectiveness is a measure of how
efficiently a computer data center uses energy; specifically, how much energy is used by the
computing equipment (in contrast to cooling and other overhead).
Quality of Service (QoS) - Quality of Service is the overall performance of a telephony or
computer network, particularly the performance seen by the users of the network.
Remote Desktop Host - A server that hosts Remote Applications as session-based
desktops. Users can access a Remote Desktop Host server by using the Remote Desktop
Connection client.
Renumbering - Network renumbering, the exercise of renumbering a network consists of
changing the IP host addresses, and perhaps the network mask, of each device within the
network that has an address associated with it.

Rollback - In database technologies, a rollback is an operation which returns the database
to some previous state. Rollbacks are important for database integrity, because they mean
that the database can be restored to a clean copy even after erroneous operations are
performed.
Remote Procedure Call (RPC) - A powerful technique for constructing distributed, clientserver based applications. The communicating processes may be on the same system, or
they may be on different systems with a network connecting them.
Recovery Point Objective (RPO) - A recovery point objective is defined by business
continuity planning. It is the maximum tolerable period in which data might be lost from an
IT service due to a major incident. The RPO gives systems designers a limit to work to.
Recovery Time Objective (RTO) - The recovery time objective is the duration of time and a
service level within which a business process must be restored after a disaster (or
disruption) in order to avoid unacceptable consequences associated with a break in business
continuity.
Software Development Kit (SDK) - A software development kit is typically a set of
software development tools that allows for the creation of applications for a certain
software package, software framework, hardware platform, computer system, video game
console, operating system, or similar development platform.
Service Level Agreement (SLA) - A service-level agreement is a part of a service
contract[disambiguation needed] where a service is formally defined. In practice, the term
SLA is sometimes used to refer to the contracted delivery time (of the service or
performance).
Software Development Lifecycle (SDLC) - Software development lifecycle - A software
development process, also known as a software development life-cycle (SDLC), is a
structure imposed on the development of a software product.
Top of Rack Switch (ToR Switch) - A Top of the Rack or (TOR) switch is a small port count
switch that sits on the very top or near the top of a Telco rack you see in Datacenters.
Traffic Shaping - Traffic shaping (also known as "packet shaping") is a computer network
traffic management technique which delays some or all datagrams to bring them into
compliance with a desired traffic profile. Traffic shaping is a form of rate limiting.
Tunneling - Computer networks use a tunneling protocol when one network protocol
(the delivery protocol) encapsulates a different payload protocol. By using tunneling one can
(for example) carry a payload over an incompatible delivery-network, or provide a secure
path through an untrusted network.
Virtual Desktop Infrastructure (VDI) - Virtual Desktop Infrastructure is a desktop-centric
service that hosts user desktop environments on remote servers, which are accessed over a
network using a remote display protocol. A connection brokering service is used to connect

users to their assigned desktop sessions.
Virtual Local Area Networks (VLAN) - In computer networking, a single layer-2 network
may be partitioned to create multiple distinct broadcast domains, which are mutually
isolated so that packets can only pass between them via one or more routers; such a
domain is referred to as a virtual local area network, virtual LAN or VLAN.
Voice over Internet Protocol (VoIP) - Voice-over-Internet Protocol (VoIP) is a
methodology and group of technologies for the delivery of voice communications and
multimedia sessions over Internet Protocol (IP) networks, such as the Internet.
Virtual Router Redundancy Protocol (VRRP) - The Virtual Router Redundancy Protocol
(VRRP) is a computer networking protocol that provides for automatic assignment of
available Internet Protocol (IP) routers to participating hosts. This increases the availability
and reliability of routing paths via automatic default gateway selections on an IP subnetwork.
VXLAN Tunnel Endpoint (VTEP) - VXLAN Tunnel Endpoint - Used for frame
encapsulation. VTEP functionality can be implemented in software such as a virtual switch
or in the form a physical switch.
Virtual Extensible Local Area Network (VXLAN) - Virtual Extensible LAN is a network
virtualization technology that attempts to ameliorate the scalability problems associated
with large cloud computing deployments. It uses a VLAN-like encapsulation technique to
encapsulate MAC-based OSI layer 2 Ethernet frames within layer 3 UDP packets.
Wide Area Network (WAN) - A wide area network is a network that covers a broad area
using leased or private telecommunication lines.
Xen - Xen is a hypervisor using a microkernel design, providing services that allow
multiple computer operating systems to execute on the same computer hardware
concurrently.